Nber Working Paper Series Market Expanding Or

NBER WORKING PAPER SERIES

MARKET EXPANDING OR MARKET STEALING? COMPETITION WITH NETWORK EFFECTS IN BIKE-SHARING

Guangyu Cao Ginger Zhe Jin Xi Weng Li-An Zhou

Working Paper 24938 http://www.nber.org/papers/w24938

NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 August 2018, Revised November 2018

Previously circulated as "Market Expanding or Market Stealing? Platform Competition in Bike- Sharing." We thank ofo executives for sharing data with us and for tirelessly answering our questions. All data employed in this article are aggregated operating indices without identifiable information on any individual user or any individual behavior. We are only authorized to use these data and the ownership of the data is retained by ofo. Comments from Marc Rysman, Feng Zhu, Joe Wyer, and seminar attendants of Peking University are appreciated. All errors are ours. Guangyu Cao personally thanks the fund support for his PhD study from Guanghua-ofo Center for Sharing Economy Research and Will Hunting Capital. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.

© 2018 by Guangyu Cao, Ginger Zhe Jin, Xi Weng, and Li-An Zhou. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source. Market Expanding or Market Stealing? Competition with Network Effects in Bike-Sharing Guangyu Cao, Ginger Zhe Jin, Xi Weng, and Li-An Zhou NBER Working Paper No. 24938 August 2018, Revised November 2018 JEL No. D22,L1,L4,L9,R4

ABSTRACT

The recent rise of dockless bike-sharing is dominated by two firms in China: one started first in 82 cities, 59 of which were subsequently entered by the second firm. Using these variations, we study how the entrant affects the incumbent's market performance. To our surprise, the entry expands the market for the incumbent. Not only does the entry boost its total number of trips and encourage more bike investment, it but also allows the incumbent to achieve higher revenue per trip, improve bike utilization rate, and form a wider and more evenly distributed network. The market expansion effect on new users dominates a significant market-stealing effect on the incumbent's old users. These findings, together with a theoretical model that highlights consumer search and network effects, suggest that a market with positive network effects is not necessarily winner-takes-all, especially when users multi-home across compatible networks.

Guangyu Cao Xi Weng Guanghua School of Management Room 304, Guanghua New Bldg Guanghua-ofo Center for Peking University Sharing Economy Resarch Beijing, China Peking University [email protected] Beijing 100871 China Li-An Zhou [email protected] Guanghua School of Management Peking University Ginger Zhe Jin Beijing 100871 University of Maryland CHINA Department of Economics [email protected] 3115F Tydings Hall College Park, MD 20742-7211 and NBER [email protected] 1 Introduction

From telephone to online platforms, many markets feature direct or indirect network effects.1 It is of concern that positive network effects could lead to winner-takes-all, where the incumbent dominates the market and competing firms find it difficult to enter and survive.2 In the meantime, multi-homing and compatibility could play an important role alleviating the anti- competitive concerns.3 How do firms compete when they face positive network effects and multi-homing users? To what extent does the entry of a competitor expand or steal the user base of the incumbent? How do price, sales, and investment of the incumbent change as a result of entry? Are there other competitive considerations besides the potential of winner-takes-all? We take these questions to the dockless bike-sharing market of China. ofo4, the first bike- sharing firm in China, was founded in 2015 by a graduate student of Peking University (PKU). Due to travel inconvenience on a large college campus, ofo started as a two-sided platform that allowed students to share privately owned bikes on campus via an online app. Soon after, the online-to-offline (O2O) platform decided to supply the GPS-tracked dockless bikes itself and effectively became one-sided. As documented by a burgeoning literature5, bike-sharing solves the “last-mile” problem of local transportation. There are positive network effects among bike riders because a user who rides a bike from A to B makes the bike available for the next rider at point B. This feature, referred to as “consumption-as-supply,” is particularly attractive in dockless bike-sharing. It no longer requires fixed docks at the origin and destination of a trip, which mitigates the potential

1Telephone is a classical example of direct network effects: consumers are more willing to join a network if they can reach more people in that network. Many two-sided platforms feature indirect network effects. For example, sellers (buyers) are more willing to join eBay if eBay attracts more buyers (sellers) on the other side; and firms are more willing to list in a yellow page if the yellow page can reach a larger number of consumers. 2The main concern is that users may be reluctant to switch away from the incumbent because they all enjoy the presence of other users in the same network. In some circumstances, users may coordinate on the wrong (inferior) network, the incumbent firm may have incentives to develop a proprietary network to lock in users, and the “excess inertia” may result in winner-takes-all. Even if multiple firms can compete to be the “winner” of the market, such competition can be inefficient from the social planner’s point of view (see the review of Farrell and Klemperer (2007)). 3On multi-homing, Caillaud and Jullien (2003) and Halaburda and Yehezkel (2013) show that multi- homing tends to intensify competition. But the recent work of Bryan and Gans (2018) suggests that competition equilibrium depends on whether multi-homing occurs on one or two sides of ridesharing platforms. On compatibility, Katz and Shapiro (1985) show that large, reputable firms tend to choose incompatibility while small, weak firms tend to choose compatibility. Farrell and Saloner (1986) further show “excess inertia” and “excess momentum” in a dynamic setting. 4“ofo” is the trademark of the firm, symbolizing a person riding a bicycle. To keep the full meaning of the trademark, we do not capitalize the first letter even if a sentence starts with “ofo.” 5Kabra, Belavina and Girotra (2016), Zheng et al. (2018) and O’Mahony and Shmoys (2015) have studied docked bike-sharing in London, New York and Paris. Pan et al. (2018) study the dockless bike- sharing firm, Mobike, in China. All of them focus on the operation of a single bike-sharing network, such as network effects, consumer demand for bikes in the existing bike network, the optimal way to locate bike docks, and algorithms that could reduce the imbalance between bike demand and bike supply.

1 imbalance of demand and supply in different locations at different times.6 When thousands of users ride ofo bikes in a small area, the wide availability of ofo bikes increases the expected probability to find a bike at the needed time and location and therefore encourages more users to use bike-sharing. In addition, more users on the road motivate ofo to put more bikes on the market, which further increases each user’s willingness to use ofo.7 Thanks to these positive network effects, ofo grew exponentially from a college campus to more than 250 cities in 20 countries by January 2018.8 ofo’s growth has attracted numerous competitors, of which Mobike is the biggest rival. From the outset, ofo and Mobike were estimated to have more than 90% of the bike-sharing market in China9, making many cities a de facto monopoly or duopoly depending on when one or both of them entered the city. If both entered the city, most consumers multi-home because the two bikes are almost perfect substitutes at the same time and location, and consumers can freely choose whichever is available at the moment.10 In this sense, the two networks are compatible and users are free to multi-home. We first demonstrate the effects of Mobike’s entry on ofo, and then present a theoretical model to explore the most likely explanations for the empirical facts. More specifically, we track news reports11 and combine them with ofo’s proprietary data. This process identifies 59 cities that were first served by ofo and then joined by Mobike. We label them ofo First cities. There are another 23 ofo Alone cities and 22 Mobike First cities.12 Because ofo started half a year earlier than Mobike, it is natural to consider ofo as an incumbent and Mobike as an entrant. With this sequence in mind, we apply difference-in-differences (DID) to the sample of ofo Alone and ofo First cities, while taking Mobike’s city-specific entry as the “treatment.” 13 Simple regressions suggest that Mobike’s entry has expanded the market for ofo, driving up ofo’s trip volume by 40.8% and ofo’s average revenue per trip by 0.041 RMB. This result is robust to heterogeneous time trends, placebo test, and an instrumental variable (IV) approach to address the potential endogeneity of Mobike entry. The instrument we use is the projected Mobike’s entry date in a city, where the projection uses the timing of Mobike’s venture capital

6Dockless bike-sharing does not completely solve the imbalance problem. See more detailed discussion in Section 2. 7This is similar to the positive feedback between demand and supply on a two-sided platform, though in our case the supply side is integrated with the firm. 8See the report from i-yiou at https://www.iyiou.com/p/64688, as of January 17, 2018. 9Industry research reports from different sources (such as iResearch, TrustDada and Analysys) cross- validate this number and some even claim that this number is estimated to be larger than 95%. 10Both apps adopt Wechat Pay and Alipay, the two most widely accepted electronic payment systems in China. 11We track news reports from 9/7/2015 to 9/14/2017. 12Our sample does not cover all the 200+ cities serviced by ofo, mostly because some cities do not have complete city attribute data from the 2016 China City Statistical Yearbook. We will elaborate our sample criterion and the definitions of ofo First, ofo Alone and Mobike First cities in Section 2. 13As detailed in Section 4, our estimation uses a long list of controls including weather conditions, air quality, calendar day fixed effects, time since ofo entry and time trends specific to predetermined city attributes. Since we do not have detailed data from Mobike, it is difficult to examine how ofo entry into Mobike First cities affects the market. Later in the paper we will describe how we use ofo data on Mobike First cities for robustness check.

2 funding (8 rounds in total) and the city’s predetermined attributes such as population, geographic feature and transportation infrastructure. When we separate new and old users within ofo, we find that Mobike’s entry has reduced the percent of old users that remain active on ofo, but this market stealing effect is dominated by the expansion in new users. Unlike other markets with positive network effects, bike-sharing allows firms to directly influence the network size by bike investment. Analysis suggests that ofo has put more bikes in the ofo First markets after Mobike’s entry, above and beyond the investment it made in ofo Alone markets; and ofo’s bike utilization rate - measured by the number of trips per ofo bike per day - has increased significantly upon Mobike entry. Above all, competition seems to have motivated the incumbent to invest more and benefit the incumbent in at least three dimensions (trip volume, revenue per trip, and bike utilization). The beneficial effect on bike utilization rate is particularly interesting, because we cannot simply explain it by the incumbent responding to entry by aggressive investment. Moreover, the effects on volume, price and utilization happen despite the fact that bike-sharing features positive network effects, multi-homing and network compatibility. This raises three immediate questions: first, what mechanism leads to the overall market expansion for the incumbent? Second, why the incumbent didn’t expand the network as much by itself before the second firm entered? Third, why the incumbent chooses its post-entry investment such that the bike utilization rate is higher after the entry than before the entry? To answer these questions and to best explain our empirical results, we model bike-sharing in monopoly and duopoly respectively. In both types of market, consumers decide whether to search for a bike, given the price of each firm and the expected probability of finding a bike. In addition to price, firms also decide on bike investment, which influences the matching probability subject to investment costs. Naturally, congestion creates a negative network effect but if the matching technology exhibits increasing return to scale, it also creates positive network effects, as more bikes on the market and more consumers searching increase the matching efficiency. In light of these network effects, we then derive how firm(s) choose price and bike investment in a subgame perfect equilibrium. We only consider symmetric equilibrium in duopoly, given the similarity between ofo and Mobike. The model predicts that, under weak conditions, duopoly always leads to a higher price than monopoly but the comparison on trade volume, bike investment and bike utilization rate depends on a few factors. The first factor is positive network effects modeled as a matching technology with increasing return to scale. When there are significant increasing return in matching, duopoly competition will generate a market expansion effect that is large enough to dominate the business stealing effect. In that case, each duopolist has incentive to make greater bike investment than the monopolist, and as a result serve more trips. When the increasing return is high enough, duopoly could also feature a higher bike utilization rate than monopoly. Another factor is investment cost. If cost per bike is constant, positive network effects would motivate the monopolist to invest in infinite bikes, leaving no room for entry. If cost per bike is increasing – a realistic assumption because it requires more effort to balance and maintain a larger and more diverse network of bikes – each firm must trade off the motive to expand due to positive network effects and the cost of investment. When investment cost (per firm) is convex enough, two firms investing at the same time is more cost-efficient than a single

3 firm making all the investment. With increasing return in matching, competitor’s investment also make one’s own investment more efficient in persuading more consumers to search and improving the matching rate. The monopolist alone cannot achieve the same efficiency, because it must shoulder the full cost to get to the same scale and it might be too convex to justify the investment.14 While increasing return matching and investment cost could explain our empirical facts, the model also offers a couple of alternative explanations for market expansion. The first alternative is that entry may raise consumer awareness (by advertising for example) and therefore increase matching efficiency for any given number of bikes. The model captures this by a multiplier in the Cobb-Douglas matching function. In equilibrium, we show that utilization rate is independent of this multiplier, hence an entry-motivated increase in the multiplier could lead to growth in bike investment but it cannot explain why ofo’s bike utilization rate is higher post entry. The second alternative explanation lies in the distribution of consumer’s private cost of using alternative transportation (the outside good). If the density of this private cost is downward sloping, an increase in that slope has two opposite effects: first, the next batch of bikes (introduced by the entrant) could persuade more consumers to join bike-sharing than the existing batch invested by the monopolist, thus generating a market expansion effect; second, there are fewer consumers with high private costs when the monopolist invests in the first batch of bikes, which discourages investment. In the model, we show that the second force dominates the first force thus the monopolist tends to invest in fewer bikes if the density is more downward sloping. This prediction is different from the marginal effect of increasing return matching, which motivates us to conduct more data analysis to distinguish these two explanations. We find that variations in private cost distribution cannot explain why the boost on ofo price and trade volume is greater for the cities in which ofo has made more bike investment before Mobike’s entry, but this empirical fact can be explained by variations in the increasing return matching technology. Our model abstracts away from local geography within a market, but it is not difficult to give it a geographic interpretation. Since the monopolist tends to first target consumers that are easiest to convert, the entry – together with the entrant’s bike investment – will help persuade the next cohort of consumers who are on the margin of using bike-sharing. To the extent that the attractiveness of the outside good is geographically dispersed, the market expansion effect of the entry also implies a geographic expansion. Using the same DID framework, we find that the entry does help ofo bikes to reach more grids in the city (each grid is defined as a square kilometer), and the network of ofo-reached grids has become more evenly distributed post entry. These findings reinforce the conclusion that competition expands the market and improves the performance of the incumbent. Our work is closest to the literature of network economy. By focusing on a market with positive network effects, we show that competition expands the overall market, even though the competing goods are close substitutes. As suggested by our model, the monopolist may not ex- haust all the positive network effects by itself, because the entrant extends the overall network of bikes and the incumbent can enjoy that positive spillover at a lower cost through competitor’s investment than through its own investment. This implication departs from the typical con-

14Our model further proves that decreasing return matching could not explain higher utilization rate post entry, nor could constant or concave cost of investment.

4 cern of winner-takes-all. Moreover, the positive network effects are reinforced by compatibility between competitors. Existing theories tend to focus on the choice of (in)compatibility while assuming users value network size and setting the cost of operation independent of network size (Katz and Shapiro, 1985).15 In bike-sharing, firms take compatibility as given but choose bike investment endogenously. Because bike investment affects the matching rate, it has a direct impact on network size in equilibrium. Our work also differs from the empirical literature of network markets. Instead of estimating how the number of users affects user evaluation16, we document the impact of entry on the incumbent, with a focus on the incumbent’s strategic choices (in price and investment) as well as the incumbent’s market performance (in trade volume, utilization rate and network reach). In doing so, we identify the effect of competition from the data directly, rather than inferring it from demand estimates and supply-side assumptions. Finally, our results highlight competition with the outside good. Many theories of two- sided markets, such as Armstrong (2006) and Bryan and Gans (2018), emphasize head-to-head competition between platforms, but assume away competition between platforms and the outside good. In our context, attracting new users to search for a bike is essential to market expansion. The market expansion effect of competition is similar to what has been found in industrial agglomeration and retailer clustering, although typical agglomeration does not feature positive network effects within one or more firms.17 Because there are positive spillovers between bike- sharing firms, our work contrasts the market stealing effect documented in other network markets (Seamans and Zhu, 2014; Angelucci and Cagé,2016). The rest of the article is organized as follows. Section 2 describes the background. Section 3 summarizes the data. Section 4 describes our main econometric specifications. Section 5 reports the baseline empirical results. Section 6 presents a theoretical model that highlights consumer search and network effects. The model also clarifies how positive network effects differ from alternative explanations. Section 7 reports further data analysis in light of the model. Section 8 concludes with policy implications.

15Katz and Shapiro (1985) model compatibility through a fixed cost, but they assume the variable cost is constant (zero) regardless of network size. 16The empirical literature has estimated network effects in many settings. For one-sided markets, the literature has estimated the direct network effects of spreadsheet, ATM, compact disk player, and VCR, and studied strategic pricing and technology adoption in light of the network effects (Saloner and Shepard, 1995; Gandal, 1994; Gandal, Kende and Rob, 2000; Park, 2004). For two-sided markets, the literature has documented the indirect network effects in yellow pages, magazines, newspapers, video games, and online platforms, and examined the implication of these network effects on pricing, advertising, product positioning, and market structure (Rysman, 2004; Kaiser and Wright, 2006; Kaiser and Song, 2009; Dubé, Hitsch and Chintagunta, 2010; Kim, Prince and Qiu, 2014; Angelucci and Cagé,2016). 17Researchers have shown that competing retailers may choose to cluster at the same mall because it lowers consumer search cost (Vitorino, 2012), automobile dealers may locate near each other despite the intensified competition (Murry and Zhou, 2018), and industries may agglomerate in the same region to enjoy positive spillovers in consumers, supplies, labors, and ideas (Ellison, Glaeser and Kerr, 2010).

5 2 Background

Since the 1960s, bike-sharing systems have gone through a few generations, mostly driven by technological development in electronically-locking racks, telecommunication systems, smart- cards and fobs, mobile phone access, and on-board computers (DeMaio, 2003; DeMaio and Gifford, 2004; DeMaio, 2009). The history of bicycle ownership and usage in China is relatively long and bike-sharing systems have followed diverse development paths in different cities (Zhang et al., 2015). Traditional bike-sharing systems provide bike rental service through stations, which means that each bike is docked at a station, riders must pick up a bike from one station, and return it to this or another station within the same network. The distance between stations and origins/destinations may be far and the capacity of stations is limited, thus the coverage of traditional bike-sharing systems is often restricted. We focus on the emerging dockless bike-sharing platforms that originated in China. Users no long need to pick up bikes from docked stations, neither do they have to dock bikes at pre-set stations. They can use smart mobile phones to scan the QR code on bike smart locks and reset it after finishing the trip at any authorized area, which is well summarized by an ofo slogan “anytime and anywhere.” From the second half of 2015, the whole bike-sharing industry has gone through explosive growth, which absorbed venture investment up to 4 billion USD and accumulatively placed more than 25 million bikes in hundreds of Chinese cities. It is estimated that the boom of dockless bike-sharing has contributed 221.3 billion RMB to economic development, created more than 390,000 jobs, and led to a welfare improvement equivalent to 175.9 billion RMB in 2017 (China Academy of Information and Communications Technology, 2018). There are also environmental benefits from dockless bike-sharing, in terms of reduced petrol consumption and decreased CO2 and NOx emissions (Zhang and Mi, 2018). ofo and Mobike are two leading firms in dockless bike-sharing, both originated in China but now operating worldwide. As the first dockless bike-sharing firm, ofo was launched on September 7, 2015 in Beijing with bikes colored yellow. At the very beginning, ofo restricted its service within college campus and limited bike outflow in many cities, which offers an opportunity for the placebo test described in Section 5. The campus-specific operation strategy was eliminated on November 17, 2016 when ofo declared full embrace of city coverage. Mobike is the main competitor of ofo, which originated in Shanghai on April 22, 2016 with bikes colored orange. As of January 2018, ofo has placed dockless bikes in more than 250 cities in 20 countries. In comparison, Mobike had placed their bikes in 176 cities of 7 countries by the end of 2017. The quick growth of ofo and Mobike has encouraged entrepreneurs and angel investors to enter the market of dockless bike-sharing. Some estimates suggest that nearly 30 new bike- sharing platforms were established in 2016 alone.18 However, various industry reports conclude that ofo and Mobike account for 90% to 95% of the bike-sharing markets from the very beginning, so that the other firms are almost negligible.19 That is why we focus on the competition between

18See the report from National Business Daily: http://www.nbd.com.cn/articles/2017-01- 05/1067671.html. 19On October 25,2017, two second-tier bike-sharing ﬁrms, Youon and Hellobike, agreed to merge. On April 4, 2018, Meituan took the full control of Mobike at a price of 2.7 billion USD. These two market events may shake the market structure profoundly, whereas both happened after our sample period.

6 ofo and Mobike, especially how the new entrant (Mobike) affects the incumbent (ofo). Because we only have access to ofo data, we collect Mobike’s entry data from media reports, and cross-validate it with postings on Mobike’s Weibo home page.20 As detailed in Section 3, our sample covers the period from May 29, 2016 to September 14, 2017, and only includes the cities that ofo has entered by September 14, 2017. Within this sample of cities, if Mobike enters the city after ofo’s entry, then the city is categorized as “ofo First.” If ofo enters the city after Mobike’s entry, it is categorized as “Mobike First.” If only ofo enters, it is “ofo Alone.” In total, our sample consists of 104 cities, of which 59 are ofo First, 23 are ofo Alone, and 22 are Mobike First. In another 6 cities out of our sample, both Mobike and ofo have entered but we could not find the exact entry date of Mobike and therefore could not define the sequence of entry precisely. We also exclude Beijing from the sample because Beijing is the birthplace of ofo and ofo had experimented with many operation policies in Beijing before it started to explore other cities. Appendix Table A1 lists the names of the 104 cities in our sample. Figure 1 plots them on the map of China. A few bike-sharing studies have examined the network feature of docked bikes. Zheng et al. (2018) set up a structural demand model to estimate consumer preference for docked bikes in the London bike-sharing system, emphasizing that the consumer must plan a trip with both the origin and the destination close to a bike station. Because of this constraint, the scope and location of the station network are important for consumer demand. They demonstrate these network effects and conclude that the existing design of the station network is far from ideal. Using data from a similar bike-sharing system in Paris, Kabra, Belavina and Girotra (2016) estimate an even-more detailed demand system. They stress that both station accessibility and bike availability are important for consumer demand, where station accessibility refers to how far a consumer must walk to a nearby bike station and bike availability refers to whether a bike is available when one walks to the station. Both accessibility and availability problems can be mitigated in dockless bike-sharing, but they are not completely eliminated. When it no longer requires a dock to park the bike, there is a possibility to find a bike near one’s home or workplace. However, less constraint on parking location may also make bikes more dispersedly distributed in the city, and therefore reduces bike availability at a particular location. In this sense, consumption-as-supply becomes more important in a dockless system, as consumers rely more on other consumers to “supply” a bike in an accessible hotspot. It also changes the nature of the network effect from a fixed network of bike stations to an evolving network of bikes “floating” throughout the city. Another problem that dockless bikes can mitigate is bike rebalance. O’Mahony and Shmoys (2015) study this problem in the docked bike-sharing system of New York City. Since demand at certain stations can be highly asymmetric during rush hours, stations at the origin of popular commuting routes will quickly run out of bikes while stations near the destination of the routes will be overwhelmed by bikes without any dock to return to. O’Mahony and Shmoys (2015) design a system that uses bike trailers to rebalance the demand and supply during rush hours and uses trucks to rebalance overnight. This rebalance problem is mitigated in dockless bike- sharing, because dockless bikes no longer need physical docks to complete the trip. However, some imbalance may still exist throughout the day, for example, traffic demand throughout the

20Weibo is one of China’s biggest Twitter-like microblogging platforms operated by Sina.

7 day may reduce bike supply needed for the afternoon rush hours, rendering a shortage at the popular origin but an excess at other locations. Pan et al. (2018) propose a deep reinforcement learning framework to solve this imbalance problem and demonstrate its effectiveness based on Mobike’s transactional data. Our paper differs from all the above, as we focus on platform competition while taking the nature of network effects as given. Because dockless systems rely on consumers’ actual demand to define bike accessibility and bike availability, the two competing systems are substitutes and complements at the same time. On the one hand, if ofo and Mobike bikes are available at the same location, they are perfect substitutes. But depletion of ofo bikes can be complemented by the remaining Mobike bikes, hence having the competitor’s bikes at the same place could increase bike availability and enhance consumer willingness to use bike-sharing. On the other hand, if ofo and Mobike bikes are placed at different locations, the overall network of bike- sharing is expanded. More consumers will find bikes accessible near the origin, and their usage will increase bike availability at the destination. It can even expand the overall network to new locations. Because the two networks are substitutes and complements to each other, it takes a full model to describe how consumer search and network effects affect each firm’s pricing and investment decisions, and whether competition would have a net market expanding or market stealing effect on the incumbent.

3 Data and Sample Construction

We combine data from several resources: ofo aggregates transactional data by time and geography, a few online platforms provide data about weather and air quality, and the 2016 China City Statistical Yearbook reports city attributes. Below we ﬁrst explain each data source, and then describe our sample construction.

3.1 Transactional Data from ofo ofo has kept full records of consumer usage, including the start and end times of each trip, longitude and latitude of the origin and the destination, listing price for the ride, and the amount actually paid after coupon redemption. From the ﬁrst usage time of each physical bike, we can also calculate ofo’s bike placement in each city over time. To protect user privacy, consumer data are aggregated to grid or city level.

We start with daily trip volume qgct, deﬁned as the total number of ofo bike trips consumed in city c, day t and grid g. Grids are deﬁned according to the longitude and latitude of the origin up to two decimal places. For example, trips originating from (23.1632◦N, 113.3578◦E) and (23.1677◦N, 113.3529◦E) will be counted as trips within the same grid (23.16◦N, 113.35◦E). P Aggregating it to the city level, we have log(Qct) = log( g qgct) for city c at day t. Daily trip volume also provides an opportunity to describe the spatial distribution of bike trips. We construct two measures: one is log(#Gridsct), namely the total number of unique grids covered by (the origin) of any ofo bike trips in a city-day. This measure aims to describe the width of the spatial network of ofo bikes as realized by consumption. The second measure aims to describe how evenly the consumption is distributed in this network. In particular, we follow

8 the definition of the Gini Coefficient, whereas “inequality” refers to trip distribution among grids instead of income distribution among population. Adopting the same method as Alesina, Michalopoulos and Papaioannou (2016), we define the base as all grids that are ever covered by ofo within a city throughout our sample period. If at day t city c no trip occurs in grid g, then qgct = 0. Assuming that there are n grids in the city and g = 1 to n are indexed in the non- Pn 1 h g=1 (n+1−g)qgct i decreasing order, we define the Gini Coverage Index as Ginict = n + 1 − 2 Pn . n g=1 qgct Another way to define Ginict is conditional on the grids that ofo has already covered in the city before Mobike’s entry, which is a subset of the base used in the first version. We will report results on both measures of Ginict. Both ofo and Mobike charge consumers by trip and time spent in the trip. ofo’s listing price is 1 RMB per hour, while Mobike’s listing price is 1 RMB per 30 minutes. The two prices are essentially identical, because bike-sharing platforms position themselves as “means of transportation for the last mile” and ofo data indicates that more than 99% of the trips end in less than 30 minutes. On top of the listing price, both platforms engaged in aggressive marketing campaigns such as trip coupons, free riding day, and monthly card for 1 RMB. These campaigns led to fluctuations in price actually paid. We thus define two variables to capture the transaction price: the first is average revenue per trip (pct), which is the simple average of total amount actually paid per ride within a city-day. It is a proxy for the average transaction price per trip. Considering that many consumers can ride for free because of coupons or other marketing activities, we also compute percent of free trips (%Freect) as an alternative measure of price within a city-day.

We define Utilizationct as the trip volume of a city-day divided by the total number of ofo bikes on the market at that city-day. Because bike investment is sparse, we aggregate the number of new bikes that ofo places in city c of month m as Investmentcm, thus the regressions on bike investment are organized by city-month instead of city-day. To examine market expansion and market stealing, it is important to distinguish old and new users of ofo. If user i registers on the ofo app at day t, she is a new user on day t and becomes an old user in any day after t. From all users’ registration history, we define log(#NewUsersct) based on the total number of new users that register on ofo in that particular city-day. We also define %ActiveOldct as the percent of old users that have used any ofo bike in that city-day, and

#T rips perOldct as the ratio between the total trips initiated by old users and the total count of old users. As mentioned in Section 2, in some cities ofo started on a college campus and gradually expanded to the rest of the city. We deﬁne the dummy 1campus equal to 1 if ofo restricts its operation within the college campus and 0 otherwise.

3.2 Weather Data and Air Quality

Weather conditions and air quality have profound impacts on the choice of travel means. Long before the emergence of bike-sharing, researchers had examined the eﬀects of weather on bike use (Hanson and Hanson, 1977; Hopkinson et al., 1989; Nankervis, 1999) and explored the impact of air pollution exposure on commuting modes (Hertel et al., 2008; Chertok et al., 2004). We use a website crawler to obtain relevant data from two open-source databases. China Meteorological

9 Data Service Center (CMDSC) provides an inquiry interface for hourly data from meteorological stations, which is averaged within each calendar day and completed through co-kriging interpolation if data from some stations are missing.21 China Air Quality Online Monitoring and Analysis Platform collects historical air quality data from the Ministry of Ecology and Environ- ment and makes it available to the public. We choose Air Quality Index as the measure of air quality in a city-day.22

3.3 Predetermined City-Level Attributes

From media report and published executive interviews, we identify four groups of city attributes that may aﬀect whether a platform enters a city: (i) economic development and overall population size are the principal determinants of potential market scale; (ii) public transportation such as bus and taxi23 may complement bike-sharing; (iii) penetration of mobile Internet and smartphones are fundamental because bike-sharing relies on real-time communication among the electronic lock of the bike, the user’s mobile phone app, and the platform’s system servers; (iv) topography (e.g. steep slope) and land forms (e.g. unpaved roads) could restrict the usage of bikes, because bikes provided by the platforms are all non-automatic. To control for the ﬁrst three aspects, we collected seven city-level variables from the 2016 China City Statistical Yearbook24: log of population, GDP per capita, the number of taxis, the number of buses, road surface, the number of mobile phones, and the number of households that have access to the Internet, which are all rescaled by total population except for log population itself. To measure terrain ruggedness, we utilize Digital Elevation Model (DEM) to calculate the average gradient for each city. All these attributes are summarized in Panel B of Table 1 and hereinafter referred to as city attributes.

3.4 Sample Construction

The original data extracted from ofo spans from September 7, 2015 to September 14, 2017. We then clean the data in a few steps: first, we exclude all autonomous prefectures and administrative districts, because they are not included in the 2016 China City Statistical Yearbook. Second, we exclude the 6 cities that Mobike entered but with missing entry dates. Without a specific entry date, we cannot confirm the entry sequence of ofo and Mobike and thus cannot define the dummy of post entry, which is the core independent variable of interest and will be introduced in the next Section. Third, we exclude Beijing from the sample. Because Beijing is

21Please see Vicente-Serrano, Saz-Sánchez and Cuadrat (2003) for detailed introduction of co-kriging interpolation. 22One potential threat to this measure lies in that air quality data disclosed by China government is under suspicion of being manipulated. However, Liang et al. (2016) finds that data from the U.S. diplomatic posts and the nearby Ministry of Environmental Protection sites produced highly consistent air quality assessment in five major cities. 23Unfortunately, the 2016 China City Statistical Yearbook does not include data on subway. But all our specifications include city fixed effects, which will absorb any time-invariant effect of subway and other omitted public transportation means. 24 The 2016 China City Statistical Yearbook reports statistics by the end of 2015, thus predetermined for our sample.

10 the birthplace of ofo, ofo had experimented with its pricing and operation strategies in Beijing extensively before it entered the second city, Shanghai. Thus, Beijing is hardly comparable to any other cities. After data cleaning, we arrive at a sample of 19,631 city-day observations, which cover 104 cities from May 29, 2016 to September 14, 2017. Table 1 summarizes the sample in two panels: one for variables at the city-day level and the other for variables at the city level. We report both panels by full sample first and then by ofo First, ofo Alone and Mobike First cities. To protect ofo’s business secrets, we mask the mean of trip volume and revenue per ride in Panel A. But from Panel B, it is obvious that ofo First cities are bigger than ofo Alone cities in almost all dimensions, including population, public transportation, and mobile/internet access. ofo First cities also have higher GDP per capita, better air quality index and lower average gradient than ofo Alone cities. Mobike First cities are more similar to ofo First cities than to ofo Alone cities. These summary statistics are consistent with the facts that bike-sharing firms tend to enter bigger and more developed cities first. Such selection prompts us to pay close attention to the comparability between ofo First and ofo Alone cities. We will deal with it in the next section. We do not report summary statistics on bike investment and bike utilization rate, because ofo designates them confidential.

4 Econometric Framework

Our main specification is difference-in-differences (DID), where we define Mobike’s entry as the “treatment” in ofo First cities, and use ofo Alone cities to control for the organic growth of ofo. In principle, we could include Mobike First cities in the control group as well, and transform the comparison into monopoly-vs-duopoly as in the theoretical model. However, we do not observe Mobike’s data before ofo’s entry into the Mobike First cities, nor can we use instrumental variable to address the endogeneity of ofo entry because we do not have data for the time that ofo had not entered. For this reason, our main specification focuses on ofo First and ofo Alone cities only, and we do not include Mobike First cities until robustness check. Specifically, the baseline specification is:

0 0 Yct = αc + γt + βP ostEntryct + Xctπ + (Sc × f(t)) θ + µGc · t + ct (1) where Yct represents outcome variables such as log(Qct), pct,%F reect, and Utilizationct at city c and date t; αc and γt denote city and time fixed effects respectively; Xct denotes weather and air quality variables; Sc denotes city attributes as of 2016; and ct is the error term. It is noteworthy that γt contains two sets of time fixed effects: the first set represents calendar date fixed effects. They aim to capture nationwide shocks on specific dates, including national holiday, nationwide news about bike-sharing, and nationwide advertising campaigns initiated by any bike-sharing platform. The second set of γt captures the intrinsic growth of ofo and is therefore defined by the number of days since ofo began operation in city c. We refer to them as relative day fixed effects.

P ostEntryct is the key regressor of interest, which takes the value of one if Mobike exists in city c on date t. For ofo First cities, P ostEntryct is zero before Mobike’s entry and becomes one at and after Mobike’s entry. For ofo Alone cities, P ostEntryct is always zero. For Mobike First cities, P ostEntryct is always one. Therefore, data on Mobike First cities do not help us identify

11 changes pre- and post-entry, though they could sharpen our understanding of ofo performance when it competes against Mobike. As stated before, we only include Mobike First cities for robustness check. To address the possibility that bike-sharing may diffuse differently in different types of cities, 25 we follow Duflo (2001) to interact city attributes (Sc) with multiple functions of time (f(t)). In particular, f(t) includes: (i) a third-order polynomial function of the relative days since ofo’s entry; (ii) calendar date fixed effects, and (iii) relative day fixed effects. In addition, we also control for linear time trends specific to ofo First cities by adding the interaction between linear time trend t and a dummy variable indicating ofo First cities (Gc). DID relies on the assumption of parallel pre-treatment trends, which could be checked by a standard event-study regression (e.g., Jacobson, LaLonde and Sullivan, 1993; Autor, 2003). Specifically, we use the following equation to test pre-treatment trends:

21 X 0 0 Yct = αc + γt + λ−kAck + βP ostEntryct + Xctπ + (Sc × f(t)) θ + µGc · t + ct (2) k=2 where Ack is a set of dummies indicating that date t is k days before Mobike’s entry into city c. We pool all days more than three weeks before Mobike’s entry as k = 21, and choose the day before Mobike’s entry (i.e., k = 1) as the omitted default category. Thus, the coefficients k=21 {λ−k}k=2 test the comparability between ofo First and ofo Alone cities for every day up to 3 weeks before Mobike’s entry. If the two groups of cities are statistically comparable, λs should be jointly indistinguishable from zero. Although including time trends and allowing them to be heterogeneous by city attributes could mitigate the concern of omitted variable bias, reverse causality is still a key identification challenge. If Mobike’s entry decision is a strategic response to ofo’s performance in a specific city, the coefficient of P ostEntryct could reflect the endogenous entry decision and does not represent the causal effect of competition on ofo. To address this concern, we need an instrumental variable that is correlated with Mobike’s entry into a city but independent of ofo’s market performance in that city. We construct the instrument based on the predicted Mobike entry date, which is the date on which we predict Mobike to enter city c according to Mobike’s VC funding rounds and c’s pre-determined city attributes. In particular, we assume Mobike could enter any city since its company establishment date (November 1, 2015). Thus, the time span between November 1, 2015 and Mobike’s actual entry date into city c is the “survival time” in a typical duration model. This is well defined for every ofo First city. For ofo Alone cities, since Mobike has not entered the city by the end of our sample, we treat the survival time as censored at 683, exactly the number of days between November 1, 2015 and September 14, 2017. We then fit the survival time in a proportional hazard duration model, where the explanatory variables are predetermined city attributes, the timing and amount of the 8-round Mobike financing from venture capital, and a new variable describing the cumulative number of days since Mobike’s latest round of VC finance. From the estimates of the duration model, we then predict the median survival time for each city and add it to the starting date (November 1, 2015). This defines the predicted entry date of Mobike.

25City attributes alone will be absorbed by city ﬁxed eﬀects.

12 From the predicted entry date, we can compute a new post-entry dummy (P ostEntry\ ct) as the

IV for P ostEntryct. We argue that the predicted Mobike entry date is likely exogenous to city-specific unknowns, because city attributes are all pre-determined and Mobike’s VC funding is not driven by a particular city. More specifically, Mobike’s VC funding may depend on ofo’s nationwide performance, which is controlled by calendar date fixed effects in the main specification, but we assume it is independent of ofo’s performance in a particular city at a particular time. We will perform statistical tests on the IV when we present the baseline results. We apply the same specifications to bike investment, but at the city-month level instead of city-day. Accordingly, we redefine P ostEntry as % of days in month m that Mobike is present in city c. Weather and air quality variables are aggregated into monthly average, and the control of time fixed effects is monthly instead of daily.

5 Baseline Empirical Results

This section reports two sets of baseline results: the ﬁrst set is on trip volume, revenue per ride, bike investment and bike utilization rate, including results with instrument and robustness checks. The second unpacks market stealing and market expanding eﬀects by new and old users.

5.1 Baseline Results

Following Equation (1), Table 2 reports the baseline DID results, where the key dependent variables are total trip volume (log(Qct)), revenue per ride (pct), and percent of free trips (%F reect).

For each dependent variable, we report the coefficient of P ostEntryct from a series of OLS regressions. The simplest one includes only city and time fixed effects (Column 1), the middle ones add interactions between f(t) and city attributes (Columns 2 to 4), and the most sophis- ticated ones add linear time trends specific to the ofo First group (Columns 5 to 7). All these columns convey the same message: Mobike’s entry has increased ofo’s trip volume and boosted ofo’s revenue per ride. If we take Column 7 as the preferred specification, it suggests that ofo’s trip volume goes up 40.8% after Mobike’s entry, ofo’s revenue per ride goes up by 0.041 RMB, and the percent of free trips goes down by 3.7 percentage points. These findings suggest a strong market expanding effect from Mobike’s entry. As shown in Appendix Table A2, similar results can be achieved when we drop ofo Alone cities from the sample (which effectively reduces the DID into just before-after comparison), or add Mobike First cities into the sample (which increases observations for post entry). To test the assumption of comparable pre-treatment trends, Figure 2 plots the point esti- k=21 mates of {λ−k}k=2 from Equation (2), along with the estimated 95% confidence intervals. The first three panels of Figure 2 correspond to the three key dependent variables (log(Qct), pct, and

%F reect). All these estimates are statistically indistinguishable from zero, neither do they imply any obvious trends jointly. This suggests that, after our control of observables, ofo Alone and ofo First cities follow similar trends before Mobike’s entry, although the two sets of cities diﬀer in absolute population and other attributes. We also perform a falsiﬁcation test by focusing on pre-entry data only (ofo First pre-entry plus ofo Alone data) and assuming a false entry on 1, 2,

13 ..., 7 days before the publicly announced entry date. Results are reported in Figure 3, along with the estimated 95% confidence interval. The first three panels correspond to the three dependent variables. For comparison, we also plot the baseline OLS results (Table 2 Column 7) on the very right. In short, the coefficients of false entry are all statistically insignificant from zero, which is very different from the baseline results. This suggests that our Mobike entry dates are accurate and the effects are attributable to the actual entry of Mobike. To further address the concern of endogenous entry, we use the predicted entry date to construct an IV for P ostEntryct. Table 3 first reports the first stage (Column 1) and then the

IV results for log(Qct), pct, and %F reect (Column 2 to 4). The Kleibergen-Paap F Test is over

8000, suggesting that our IV is strongly correlated with P ostEntryct. After using the IV, the key coefficients of P ostEntryct (β) have the same sign and similar magnitudes as in the OLS regressions. Table 4 reports the OLS and IV results for the effect of entry on bike utilization rate and bike investment. The utilization regressions are at the city-day level, while the investment regressions are at the city-month level. For the OLS columns, we use the specification that includes the most extensive set of controls, as in Column 7 of Table 2. For the 2SLS columns with IV, we use the same instrument as before, except that the instrument is aggregated into a monthly average, i.e. % of days in month m that we predict Mobike to be present at city c. Both OLS and 2SLS results suggest that Mobike’s entry have motivated ofo to place more bikes in the city and enjoy a significant boost in bike utilization. We perform two robustness checks on the IV results in Appendix Table A3. First, since the proportional hazard model relies on the functional form of baseline hazard, we confirm that results are stable when we use Weibull (reported), log-normal, or log-logistic distribution for baseline hazard. Second, Mobike was established on November 1, 2015 but did not enter the first city (Shanghai) until April 22, 2016. We have tried to use December 1, 2015, January 1, 2016, February 1, 2016, March 1, 2016 and April 1, 2016 as alternative starting dates. Results under these alternatives are similar to what is reported in Table 3 and Table4, except that the results on bike investment lose statistical significance at the 90% level if we assume the baseline hazard distribution is log-logistic or lognormal. One possible explanation is that investment is sparse and therefore sensitive to functional form. However, even in these marginal results, the coefficient of P ostentryct has the same sign and similar magnitude as in the baseline result. Above all, we find that Mobike’s entry has increased ofo’s trip volume and revenue per trip, has encouraged the incumbent to place more new bikes on the market, and has helped ofo to enjoy a higher bike utilization rate. Robustness checks further suggest that these effects are unlikely driven by omitted variable bias or endogenous entry.

5.2 New and Old Users

If entry has led to an increase in price and trade volume at the same time, it suggests market expansion. However, since Mobike and ofo bikes are almost perfect substitutes at the same time and location, the entry could have a market stealing eﬀect as well. We examine this possibility by separating new and old users within ofo. Note that both new and old are from ofo’s point of view, as we do not know whether a user has also downloaded the Mobike app or not.

14 Results are presented in Table 5. The OLS results suggest that Mobike’s entry has increased the number of new users (for ofo) by 65.2%, and this effect is even greater if we use the instrument (73.5%). However, percent of active old users declines 4.1-4.4 percentage points post entry, which is a significant fraction of the sample mean26. Because every new user becomes an old user after the registration day, the pool of old users is cumulative over time. Thus 4.1-4.4% of this pool is a significant market stealing effect if all of them switch to Mobike. Conditional on old users, Columns 5-6 show that the average number of trips they take on ofo does not change significantly post Mobike entry. As shown in Columns 7-10, Mobike’s entry has increased revenue per trip for both new and old users, at a similar magnitude. In short, we observe market expansion into new users and market stealing of old users, the sum of which gives rise to the overall market expansion effects documented in the baseline results. To summarize, Mobike’s entry has created a net market expansion for ofo, despite some market stealing effects on old users. A potential explanation is that Mobike’s marketing campaign, including the sight of orange bikes on the road, may have motivated more consumers to use bike-sharing. Both Mobike and ofo have issued coupons to lure new users, which could contribute to market expansion as well. However, the entry has a significant, positive effect on price, suggesting that the market expansion is not solely driven by the price competition between the two firms. What is the mechanism behind such price-boosting market expansion? Why does ofo find it worthwhile to put even more bikes on the market after Mobike has entered the market with many orange bikes? If bike investment is just another form of marketing, how could bike utilization rate increase post entry? And if the higher bike utilization rate implies that bike investment is effective in attracting more users, why didn’t ofo make the extra bike investment until Mobike entered? We attempt to answer these questions in a theoretical model.

6 Theoretical Model

In this section, we develop a theoretical model to explain the empirical findings presented in the previous section. Our goal is to use the simplest model to explain how entry could generate higher price, higher trade volume, higher investment and higher utilization rate in a unified framework. It is not difficult to come up with a story to explain higher price and higher trade volume, or even higher investment. For example, Mobike’s marketing campaign can be one of the potential explanations. But the real challenge is to explain why bike utilization rate is higher post entry. As detailed below, this requires a particular combination of consumer search, network effects and investment cost.

6.1 Model Setup

We consider a market consisting of measure 1 of consumers. Each consumer has to ﬁnish a trip with value normalized to 1. A consumer can use either bike-sharing or his own way to ﬁnish this trip (the outside good). If he uses his own way such as buying his own bike, driving his private car or calling a taxi, the private cost is distributed on the interval [0, 1] according to the distribution function F (c) = 1 − (1 − c)θ where θ > 0 is common knowledge. Obviously,

26We are not allowed to report the sample mean because it is a business secret.

15 when θ = 1, this distribution function is the same as the uniform distribution. When θ > 1, the density function is decreasing in c; while when θ < 1, the density function is increasing in c. If a consumer uses bike-sharing, he pays a price for the service. However, the consumer may not find a bike nearby. We assume that the probability of finding a bike is determined by an aggregate matching function. In particular, if there are measure u of consumers searching for a bike and there are measure v of bikes in the market, then the total measure of matches is given by m(u, v) = Avαuβ with 1 ≥ α, β > 0. This Cobb-Douglass matching function is widely used in the literature. It is usually assumed that the matching function exhibits constant return to scale: α + β = 1. But we do not impose this assumption here, and also allow decreasing or increasing return to scale α + β > 1.27 An increasing return matching technology could reflect the positive network effect in the bike-sharing industry: the consumers are actually transporting bikes for the firm when they are using the service of bike-sharing. Therefore, as more consumers are sharing bikes, it is more likely for other consumers to find a bike. In comparison, the multiplier A is another parameter that governs matching efficiency. Since A is a constant independent of the number of bikes (v) and the numbers of searching consumers (u), it captures network-neutral technology factors such as consumer awareness of bike-sharing and the quality of bike-sharing apps. Under the above matching function, the probability for a consumer to find a bike is given m(u,v) α β−1 by q = u = Av u . We assume that a consumer only search once. If he could not find a bike, then he receives an outside value of 0. This assumption reflects the fact the bike-sharing aims to solve the “last mile” problem for the consumer. If a consumer cannot find a bike for the trip, he usually will try other ways of transportation instead of keeping searching bikes.28 We will consider two cases. In the first case, there is one monopolist operating in the market; while in the second case, there are two duopoly firms competing in the market. In both cases, the sequence of move is that first, the firms set up the prices and total measures of bikes put into the market; and then the consumers then choose between bike-sharing and their own ways to finish the task. If a consumer chooses bike-sharing, he finds a bike with some probability. If he finds a bike, he will use it as long as the price charged is lower than 1. Otherwise, he will take the outside option. In the first monopoly case, it is the monopoly firm who sets the price p and the total measure 1 γ of bikes v put into the market. We assume that the investment cost function is ψ(v) = γ φv with γ > 0 capturing the concavity/convexity of the cost function. In the second duopoly case, we assume that firm 1 is the same as the monopoly firm while firm 2 is a new entrant with the same cost function ψ(v).29 In this case, the two duopoly firms simultaneously choose the measures of

27Such an assumption is also adapted in many other studies, e.g., Gan and Zhang (2006), Petrongolo and Pissarides (2006), Gavazza (2011), Bleakley and Lin (2012). 28In reality, a consumer that could not ﬁnd a bike may still use alternative transportation to complete the trip, but there is a delay as compared to using the alternative transportation at the very beginning. For example, one may get to work on time if she calls a taxi at time t or searches for a bike at time t and rides the bike at time t + 1. However, she will be late for work if she calls a taxi at time t + 1. Our assumption on the value of ride sharing and alternative transportation (before search) is just a normalization. In the above example, it is equivalent to assuming the value of getting late to work is 0, the value of biking to work on time is 1, and the value of calling taxi to work on time is 1 − c. 29In general, there is no need to assume that the ﬁrms have the same cost function. But the symmetric

16 bikes put into the market (v1, v2) and the prices (p1, p2). Given v1 and v2, the probability for a consumer to find a firm 1’s bike is given by q = A(v + v )αuβ−1 v1 , where A(v + v )αuβ−1 1 1 2 v1+v2 1 2 is the probability of finding a bike and v1 is the probability that this bike belongs to firm 1. v1+v2 The underlining assumption is that search is purely random and the consumers cannot target which bike to search. Similarly, the probability for a consumer to find a firm 2’s bike is given by q = A(v + v )αuβ−1 v2 . 2 1 2 v1+v2 We aim to solve the subgame perfect equilibria of this model. In the first monopoly case, the key is to solve the monopoly price pm and bike investment vm in equilibrium, while in the second d d duopoly case, the key is to solve the price and investment of firm 1 p1 and v1 in equilibrium. m m d d The final step is to investigate the impact of firm 2’s entry by comparing (p , v ) with (p1, v1).

6.2 Equilibrium Analysis

6.2.1 Monopoly Case

We solve the subgame perfect equilibrium in the monopoly case by backward induction. Given p and v, a consumer will choose bike-sharing if

q∗(1 − p) ≥ 1 − c, or c ≥ 1 − q∗(1 − p), where q∗ is the equilibrium probability that a consumer ﬁnds a bike. Hence, under the distribu- tional assumption of c, the total measure of searching consumers is given by u = (q∗(1 − p))θ. m(u,v) α β−1 ∗ This together with the condition q = u = Av u pins down q :

q∗ = Avα((q∗(1 − p))θ)β−1, which implies 1 α θ(β−1) q∗ = A 1+θ(1−β) v 1+θ(1−β) (1 − p) 1+θ(1−β) .

When the monopolist sets the price and total measure of bikes, the objective is to maximize:

Avα(q∗(1 − p))θβp − ψ(v), where Avα(q∗(1 − p))θβ is the total measure of matches and for each match, the monopolist receives a revenue of p. Plugging the expression of q∗ into the above maximization problem yields 1+θ α(1+θ) θβ A 1+θ(1−β) v 1+θ(1−β) (1 − p) 1+θ(1−β) p − ψ(v).

Clearly, on the one hand, by putting more bikes into the market, the total revenue will increase by generating more matches, but the cost also increases. On the other hand, by setting a higher price, the revenue for each match will increase, but the total measure of matches also decreases as fewer consumers choose bike-sharing. Based on this tradeoﬀ, the monopoly chooses case is easy to solve theoretically. Moreover, in reality, the two leading bike-sharing ﬁrms, ofo and Mobike, are quite symmetric.

17 the amount of total investment and price, vm and pm are solved from the ﬁrst order conditions. First of all, the monopoly price is given by θβ pm = 1 − pm, 1 + θ(1 − β) which implies that 1 + θ(1 − β) pm = < 1. 1 + θ α(1+θ) Second, when 1+θ(1−β) < γ, the optimization problem is concave in v, and hence the monopoly investment vm satisﬁes

α(1 + θ) 1+θ α(1+θ) −1 θβ A 1+θ(1−β) v 1+θ(1−β) (1 − pm) 1+θ(1−β) pm = φvγ−1, 1 + θ(1 − β) which implies

1  1+θ θβ  α(1+θ) γ− α(1 + θ) A 1+θ(1−β) (1 − pm) 1+θ(1−β) pm 1+θ(1−β) vm = . 1 + θ(1 − β) φ 

The above result can be summarized by the following lemma:

α(1+θ) Lemma 6.1 Assume that 1+θ(1−β) < γ. Then, there exists a unique subgame perfect equilibrium in the monopoly case. In this equilibrium, the price is 1 + θ(1 − β) pm = (3) 1 + θ and the investment is

1  1+θ θβ  α(1+θ) γ− α(1 + θ) A 1+θ(1−β) (1 − pm) 1+θ(1−β) pm 1+θ(1−β) vm = . (4) 1 + θ(1 − β) φ 

α(1+θ) Notice that the condition 1+θ(1−β) < γ does not rule out the possibility of γ < 1. For example, if α + β = 1, there exists γ < 1 satisfying the above condition. Moreover, for γ ≥ 1, the above condition is satisﬁed when γ − α + θ(γ − 1)(1 − β) α + β < 1 + . θ

α(1+θ) So there exists α + β > 1 satisfying 1+θ(1−β) < γ.

6.2.2 Duopoly Case

Now we move to the duopoly case. In this case, we also first solve the consumer’s problem ∗ of whether to choose bike-sharing or not. Denote q1 to be the equilibrium probability that a ∗ consumer finds bike 1, and q2 to be the equilibrium probability that a consumer finds bike 2. Recall that given v and v , these two probabilities are given by q = A(v + v )αuβ−1 v1 and 1 2 1 1 2 v1+v2 q = A(v + v )αuβ−1 v2 . 2 1 2 v1+v2 A consumer will choose bike-sharing if

∗ ∗ q1(1 − p1) + q2(1 − p2) ≥ 1 − c,

18 or ∗ ∗ c ≥ 1 − q1(1 − p1) − q2(1 − p2). Similar to the monopoly case, we can solve the equilibrium probabilities as:

∗ α ∗ ∗ θ(β−1) v1 q1 = A(v1 + v2) (q1(1 − p1) + q2(1 − p2)) v1 + v2 and ∗ α ∗ ∗ θ(β−1) v2 q2 = A(v1 + v2) (q1(1 − p1) + q2(1 − p2)) . v1 + v2 Firm 1’s proﬁt then can be written as:

α v1 ∗ ∗ θβ A(v1 + v2) (q1(1 − p1) + q2(1 − p2)) p1 − ψ(v1). v1 + v2 Compared with the proﬁt in the monopoly case

Avα(q∗(1 − p))θβp − ψ(v), we can observe two opposing effects. The first business stealing effect comes from the observation that (v + v )α v1 < vα for any v > 0. Basically, the operation of firm 2 decreases firm 1’s 1 2 v1+v2 1 2 total measure of matches for a given v1, because some of the consumers are stolen by firm 2. ∗ ∗ The second market expansion effect comes from the term q1(1 − p1) + q2(1 − p2). The existence of firm 2 attracts more consumers into the market, and hence increases the total measure of matches for firm 1. ∗ ∗ We can first solve q1 and q2 as functions of v1, v2, p1, p2. Plugging these functions into the firms’ profits yields:

1+θ 1+θ (1+θ)(1−α) v θβ 1+θ(1−β) 1+θ(1−β) − 1+θ(1−β) 2 1+θ(1−β) π1 = A v1 (v1 + v2) [(1 − p1) + (1 − p2)] p1 − ψ(v1), v1 and 1+θ 1+θ (1+θ)(1−α) v θβ 1+θ(1−β) 1+θ(1−β) − 1+θ(1−β) 1 1+θ(1−β) π2 = A v2 (v1 + v2) [(1 − p2) + (1 − p1)] p2 − ψ(v2). v2 d d d d The firms simultaneously choose (v1, p1) and (v2, p2) to maximize profits. We will focus on d d d d the symmetric equilibrium: v1 = v2 and p1 = p2. First, the first order condition with respect to p1 is:

θβ v2 p1 = (1 − p1) + (1 − p2). 1 + θ(1 − β) v1 In the symmetric equilibrium, it is straightforward to derive

d d d 2 p1 = p2 = p = θβ . 2 + 1+θ(1−β)

Second, we can take first order conditions with respect to v1 in the firm 1’s profit function. d d d d The equilibrium investments v1 and v2 can be solved by plugging p1 and p2 into the first order d d d d conditions. In the symmetric equilibrium, we obtain v1 = v2 = v , with v satisfying: 1+θ 1 + θ (1 + θ)(1 − α) θβ γ− α(1+θ) A 1+θ(1−β) − − ω = φv 1+θ(1−β) , 1 + θ(1 − β) 2[1 + θ(1 − β)] 2[1 + θ(1 − β)] where θβ−(1+θ)(1−α) θβ ω = 2 1+θ(1−β) (1 − pd) 1+θ(1−β) pd. The above result can be summarized by the following lemma:

19 α(1+θ) Lemma 6.2 Assume that 1+θ(1−β) < γ. Then, there exists a unique symmetric subgame perfect equilibrium in the duopoly case. In this equilibrium, both ﬁrms set prices to be

d 2 p = θβ (5) 2 + 1+θ(1−β) and investments to be

1 α(1+θ) 1+θ 1 + θ (1 + θ)(1 − α) θβ ω γ− vd = A 1+θ(1−β) − − 1+θ(1−β) . (6) 1 + θ(1 − β) 2[1 + θ(1 − β)] 2[1 + θ(1 − β)] φ

6.3 Equilibrium Comparison

In this section, we compare the unique subgame perfect equilibrium in the monopoly case with the unique symmetric subgame perfect equilibrium in the duopoly case. We aim to compare prices, investments, trade volumes, and utilization rates in different cases. Trade volume is defined as the total measure of successful matches for firm 1. In the monopoly case, it is d m m m d d d d v1 TV = m(u , v ), while in the duopoly case, it is TV = m(u , v1 + v2) d d . Utilization rate v1 +v2 m TV m is defined as the average number of match per bike. It is r = vm in the monopoly case, and d TV d r = d in the duopoly case. v1 Our first observation is the following:

Proposition 1 In equilibrium, it is always the case that pd > pm.

Proof From equations (3) and (5), we obtain 2 1 pd = > pm = . 2 + θβ 1 + θβ 1+θ(1−β) 1+θ(1−β)

The above proposition claims that the duopoly price is always higher than the monopoly price. This result is very intuitive. When there is only one firm in the market, raising price will reduce the number of searchers and this negative impact is fully incorporated in the monopolist’s pricing decision. If the prevailing price is already optimal for the monopolist, the marginal benefit of the price hike (higher profit per match) is equal to the marginal cost of the hike (fewer consumers searching). In contrast, if the firm faces competition in duopoly, its price hike will affect the number of searchers as before but this hurts both firms. Since each firm does not incorporate the negative externality its price hike imposes on the competitor, competition reduces the marginal cost of price hike, while the marginal benefit remains the same. In other words, competition blunts the negative impact of price hike for each firm’s individual demand. This is equivalent to reducing the demand elasticity facing each firm, creating an extra incentive to raise price. For this reason, the model shows that price increases when the market moves from monopoly to duopoly, regardless of the shape of the private cost distribution or the efficiency of the matching technology. Note that we are not the first one finding price increase with competition. In a model of price search, Stahl (1989) shows that the equilibrium price approaches monopoly price when the number of firms increase, because the probability of finding the lowest price decreases exponentially with the number of firms. In a market where each seller faces loyal and switching consumers,

20 Rosenthal (1980) shows that competition may reduce each seller’s share of the switching group and therefore incentivize it to charge higher price among the remaining loyal consumers. Al- lowing product differentiation, Chen and Riordan (2008) show that price may increase from monopoly to duopoly, if consumer preferences for the two products are sufficiently diverse and negatively correlated. Our model differs from all of them, because we assume consumers only search once (per episode) and therefore price shopping does not occur within the bike search. Our next results focus on the comparison of the utilization rates rd and rm. By definition, we obtain 1+θ α(1+θ) θβ A 1+θ(1−β) (vm) 1+θ(1−β) (1 − pm) 1+θ(1−β) rm = . vm m α(1+θ) m m m γ−1 And the first order condition with respect to v implies that 1+θ(1−β) r p = φ(v ) , which implies φ(vm)γ−1 rm = . (7) α(1+θ) m 1+θ(1−β) p Similarly, we derive φ(vd)γ−1 rd = . (8) h 1+θ (1+θ)(1−α) θβ i d 1+θ(1−β) − 2[1+θ(1−β)] − 2[1+θ(1−β)] p The above equations implies that the comparison of rd and rm does not depend on A. Hence, we can even allow A to increase in the duopoly market and get the same result. The increase of A could occur if the entry of the second firm makes more consumers aware of bike-sharing or leads to a higher consumption value to the consumers (e.g., Mobike’s marketing campaign).30 Although the increase in A will lead to a higher investment, it cannot lead to a higher utilization rate. To obtain higher utilization rate in equilibrium, we need the following condition:

Proposition 2 It is possible to have both higher investment vd > vm and higher utilization rate rd > rm in equilibrium only when γ > 1. And when γ > 1, rd > rm is satisﬁed only if vd > vm.

Proof From equations 7 and 8, we obtain rd > rm if

α(1+θ) pd vd < ( )γ−1 1+θ(1−β) . pm vm 1+θ (1+θ)(1−α) θβ 1+θ(1−β) − 2[1+θ(1−β)] − 2[1+θ(1−β)] Rearranging terms implies that vd α(1 + θ) + (1 + θ) − θβ ( )γ−1 > > 1. vm α[2(1 + θ) − θβ] Therefore, if vd > vm, the above inequality can be satisﬁed only when γ > 1. Moreover, when d m d m γ > 1, in order to get r > r , we must have v > v .

The above proposition implies that higher investment and higher utilization rate can both exist only when the investment cost function is convex. So we will make the assumption that γ > 1 in the subsequent analysis. Since a higher investment is a prerequisite for a higher utilization rate, our next proposition investigates when the equilibrium investments satisfy vd > vm.

30For example, consider the case that the total measure of consumers increase from 1 to η > 1 due to the fact that more consumder are aware of bike-sharing. This is equivalent to an increase in the matching eﬃciency increases from A to Aηβ > A.

21 Proposition 3 The equilibrium investments satisfy vd > vm if

θβ θβ−(1+θ)(1−α) 2α(1 + θ) (1 − pm) 1+θ(1−β) pm 2 1+θ(1−β) > . (9) α(1 + θ) + 1 + θ(1 − β) θβ (1 − pd) 1+θ(1−β) pd

Moreover, when θβ − (1 + θ)(1 − α) ≤ 0, we must have rd < rm.

Proof Condition(9) directly comes from equation (4) and (6). vd > vm if

1 + θ (1 + θ)(1 − α) θβ α(1 + θ) θβ − − ω > (1 − pm) 1+θ(1−β) pm, 1 + θ(1 − β) 2[1 + θ(1 − β)] 2[1 + θ(1 − β)] 1 + θ(1 − β) which implies condition (9). From equations 7 and 8, we know that the comparison of rd and m vd γ−1 r depends on the comparison of ( vm ) and

1+θ (1+θ)(1−α) θβ pd − − α(1 + θ) + (1 + θ) − θβ Ω 1+θ(1−β) 2[1+θ(1−β)] 2[1+θ(1−β)] = > 1. , pm α(1+θ) α[2(1 + θ) − θβ] 1+θ(1−β)

From equation (4) and (6), we derive

γ−1 d d α(1+θ) v θβ−(1+θ)(1−α) 1 − p θβ γ− ( )γ−1 = 2 1+θ(1−β) ( ) 1+θ(1−β) Ω 1+θ(1−β) . vm 1 − pm

vd γ−1 Therefore, the comparison of ( vm ) and Ω is equivalent to the comparison of

γ−1 d α(1+θ) θβ−(1+θ)(1−α) 1 − p θβ γ− Γ 2 1+θ(1−β) ( ) 1+θ(1−β) 1+θ(1−β) , 1 − pm and γ−1 1− α(1+θ) γ− Ω 1+θ(1−β) .

Notice that whenever θβ − (1 + θ)(1 − α) ≤ 0, we must have

θβ−(1+θ)(1−α) 2 1+θ(1−β) ≤ 1 and α(1 + θ) θβ − (1 + θ)(1 − α) = 1 + ≤ 1. 1 + θ(1 − β) 1 + θ(1 − β) So it is straightforward to see that Γ < 1 since we have 1 − pd < 1 − pm from Proposition 1, and

γ−1 1− α(1+θ) γ− Ω 1+θ(1−β) ≥ 1

γ−1 since Ω > 1 and 1 − α(1+θ) ≥ 0. As a result, whenever θβ − (1 + θ)(1 − α) ≤ 0, we should γ− 1+θ(1−β) vd γ−1 d m get ( vm ) < Ω, which implies that r < r .

Condition (9) in the above proposition gives the condition under which the equilibrium θβ−(1+θ)(1−α) investments satisfy vd > vm. Intuitively, the term 2 1+θ(1−β) in condition (9) measures the joint effects of the market expansion and business stealing effects. From the expressions of the firms’ profit functions, it is easy to see that in the symmetric equilibrium, the business

22 − (1+θ)(1−α) stealing effect is proportional to 2 1+θ(1−β) while the market expansion effect is proportional to θβ θβ−(1+θ)(1−α) 2 1+θ(1−β) . So the term 2 1+θ(1−β) measures the joint effect. And this term becomes larger as the market expansion effect becomes stronger than the business stealing effect. The second part of Proposition 3 implies that when the business stealing effect dominates the market expansion θβ−(1+θ)(1−α) effect ( 1+θ(1−β) ≤ 0), it is impossible to achieve a higher utilization rate in the duopoly case. Therefore, we need to further consider the case of a sufficiently degree of increasing return 1−α to scale α + β > 1 + θ to satisfy θβ − (1 + θ)(1 − α) > 0. Notice that this condition rules out the case of constant return to scale: α + β = 1. In other words, we should always get a lower utilization rate under constant return to scale. Our last observation claims that it is possible to generate both vd > vm and rd > rm when both α + β and θ are sufficiently large.

Proposition 4 Suppose that both α and β go to one, and θ goes to inﬁnity. Then we must have both vd > vm and rd > rm.

Proof Consider the extreme case of α = β = 1. In this case, the left-hand side of condition (9) is θβ θ 2(1+θ) (1−pm) 1+θ(1−β) pm 2 while the right-hand side is (1+θ)+1 θβ . Clearly, when θ goes to infinity, the left- (1−pd) 1+θ(1−β) pd hand side goes to infinity as well while the right-hand side stays bounded. As a result, condition d m vd γ−1 α(1+θ)+(1+θ)−θβ (9) is satisfied. Moreover, from the proof of Proposition 2, r > r if ( vm ) > α[2(1+θ)−θβ] . Notice that the right-hand-side of the above inequality is one when α = 1, while the left-hand- d m d m side goes to infinity as θ goes to infinity. Therefore, we must have both v > v and r > r . Figure 4 illustrates how the parameter values α and β affect the comparison of rd and rm when we fix θ = γ = 2. We can see several interesting features from Figure 4. First of all, the dashed line in Figure 4 is an upper bound on the value of β. This comes from the requirement α(1+θ) 1+θ(1−β) < γ. Second, the region between the solid line and the dashed line represents the parameter values of α and β under which we have rd > rm. From Proposition 2, in this region we have vd > vm, which also implies TV d = vdrd > T V m = vmrm. Notice that there is a sufficient degree of increasing return to scale in this region: α + β > 1. In particular, this region exists only when α > 0.5. This is because from Proposition 3, we require θβ > (1 + θ)(1 − α). So α cannot be too low such that the above inequality is satisfied for some β ≤ 1. Finally, as α goes to one, α + β on the solid cutoff line decreases to be lower than 1.2, which implies that a small degree of increasing return to scale is enough to generate vd > vm. Although it is difficult to conduct theoretical comparative static analyses, Figures 5 and 6 numerically investigate how the change in parameters affects the equilibrium outcomes in monopoly and duopoly cases. Figure 5 compares the equilibrium outcomes in monopoly and α duopoly cases when we change parameter z = α+β while keeping the ratio β to be a constant. So in this numerical exercise, α and β changes with z at the same rate. Consistent with Proposition 1, we always have pd > pm as seen in Figure 5a. Figure 5b plots how vm and vd change with z, and shows that vm > vd when z is low and vice versa. This is consistent with our discussion of condition (9). Figures 5c and 5d show that both trade volumes and utilization rates follow similar patterns: they are higher in the duopoly case only when z is sufficiently large. This is also consistent with Figure 4: we need the degree of increasing return to scale to be sufficiently large to guarantee rd > rm, vd > vm and TV d > T V m.

23 Figure 6 compares the equilibrium outcomes in monopoly and duopoly cases when we change the distribution parameter θ. The findings are similar to the ones in Figure 5. In particular, we find that when we fix other parameters, it is also possible to have rd > rm, vd > vm and TV d > T V m when θ is sufficiently large. Moreover, although sufficiently high θ and sufficiently high z are both consistent with our baseline results, an increase in θ always decreases bike investment in the monopoly case while an increase in z always increases bike investment in the monopoly case. This is very intuitive. An increase in θ implies a smaller measure of consumers with high cost c and hence makes the market unattractive. A monopoly firm will optimally respond by lowering its bike investment. On the contrary, an increase in z implies a larger degree of increasing return to scale and hence makes the market more unattractive. A monopoly firm will optimally respond by increasing its bike investment. To summarize, we find that it is possible that investment, trade volume and utilization rate all increase in the duopoly case when θ or z are sufficiently large. A high z can be interpreted as a large enough increasing return to scale in the matching function. A high θ can be interpreted as the density of the private cost of the outside good (c) decreasing at a sufficiently high speed as c increases. In both situations, the market expansion effect is sufficiently large to dominate the business stealing effect, which implies the following testable implications:

1. After the entrant’s entry, the incumbent’s price goes up;

2. After the entrant’s entry, some of the incumbent’s old customers are stolen by the entrant while the incumbent can get new customers due to market expansion;

3. After the entrant’s entry, the incumbent’s bike investment may increase when θ is suﬃ- ciently high or z = α + β is suﬃciently high;

4. After the entrant’s entry, both the incumbent’s trade volume and utilization rate can also go up when θ is suﬃciently high or z = α + β is suﬃciently high.

6.4 Discussion of the model

The model highlights a few important features of bike-sharing: consumer search, matching technology, investment cost and the outside good. We now discuss each respectively.

6.4.1 Consumer search

Using bike-sharing requires the consumer to initiate a bike search. Assuming the search is random and once-for-all, we downplay competition within the search process. As a result, competition only affects the search result through bike investment (which affects the matching probability), and the number of consumers that decide to search (which depends on the expected matching rate, consumer’s private cost of using the outside good, and the price of each bike- sharing firm). Because of these assumptions, competition tends to reduce the demand elasticity facing each firm, which naturally leads to our Proposition 1: entry increases the equilibrium price. The assumption that search is once-for-all might seem strong at the first glance. It nonethe- less reflects an important feature at the current stage of bike sharing: the scarcity margin is

24 more important than price margin. That is to say, most consumers use bike-sharing for com- mute (rather than recreation) and they care more about finding a bike than the relatively small price differential. Therefore, they will not continue searching as in standard search models. Our model remains appropriate as long as there is a keen concern of not finding a bike in time. It is also possible to develop a more complete model incorporating search cost, but we choose not to do so for two reasons. On the one hand, the results in such a full-fledged model is likely similar to those in the current model.31 On the other hand, a very detailed model of consumer search might blur the main purpose of our model, which is to explain the bigger puzzle in higher utilization rate post entry. We do not even need a complicated search model to explain why price and investment increase after entry (e.g. Mobike’s marketing campaign alone could achieve it), but the real challenge is explaining the increase in utilization rate while keeping the changes in price and investment consistent with the facts. For this purpose, we believe our current model of consumer search provides the simplest way to illustrate the main driving force behind the data.

6.4.2 Network eﬀects in the matching technology versus investment cost

The second key feature is network effects. Given a fixed supply of bikes, the more consumers search for a bike, the less likely each consumer gets matched to a bike. This is a negative network effect of congestion. However, if bikes and searchers increase proportionally, the matching rate will increase or decrease depending on the matching technology. When the matching technology has increasing return to scale (z = α + β > 1), two million consumers searching for 2N bikes will have a higher matching rate than the first million consumers searching for N bikes, and the improved matching efficiency will encourage more consumers to join the search. This creates a positive network effect. Furthermore, because matching rate improves by scale, each bike has a higher rate of utilization as the numbers of consumers and bikes increase. If investment cost per bike does not increase as fast as the utilization rate, firm(s) has incentive to invest in more bikes. This creates a second positive feedback in the system, similar to what we have seen on two-sided platforms (e.g. more sellers attract buyers, and more buyers attracts sellers). If the network effect is positive, it could have an important impact on market structure. For instance, if the positive network effects are always large enough to swamp any increase in investment cost, the market is winner-takes-all, because the monopolist has incentive to invest in infinite bikes, leaving no room for other firms to enter. This possibility is ruled out in our model α(1+θ) when we impose the assumption 1+θ(1−β) < γ. Under this assumption, the cost of investment will eventually dominate the expanding incentive driven by the positive network effects. When the positive network effects are sufficiently large but not too large to trigger winner- takes-all, each duopolist will engage in more bike investment than the monopolist. This is because each duopolist free rides on the competitor’s investment. Every bike invested by firm 2 costs nothing to firm 1, but expands the overall market and benefits firm 1. In this sense, it is more cost-efficient than firm 1’s own investment. While free-riding incentive often exists in duopoly, it is magnified when the market enjoys large, positive network effects. With such

31For example, due to positive search cost, the famous Diamond paradox claims that ﬁrms will charge monopoly price, which is quite close to the equilibrium price in our current model.

25 network effects, competitor’s investment will make one’s own investment more efficient in persuading more consumers to search and improving the matching rate. The monopolist alone cannot achieve the same efficiency, because the monopolist must invest at its own cost to get to the same scale and its cost function might be too convex to justify the investment. In that situation, we believe there is no first mover or second mover advantage, though the model has not addressed sequential entry explicitly. Since monopoly profit is greater than zero and duopoly profit is greater than monopoly profit, entry is always preferred to no entry. In comparison, when the positive network effects are relatively small, the model predicts that each duopolist invests less than the monopolist. If all past investments are sunk, the monopolist may find it over-invested when the second firm enters. As a result, the entrant enjoys extra free ride from the monopolist’s over-investment, which implies a first-mover disadvantage and a second-mover advantage. It is difficult to test whether this strategic concern applies to our data, but we have addressed the potential endogeneity of Mobike entry in the empirical analysis. The fact that we find ofo invests more after Mobike’s entry suggests that this regrettable situation (for the monopolist) is unlikely in our setting.

6.4.3 Competition with the outside good

The extent to which competition expands the market also depends on the outside good, namely the distribution of consumers’ private cost of using the outside good (c). The higher the private cost, the more attractive is bike-sharing relative to the outside good (private bike, private car, taxi, bus, etc.). Since the monopolist must first attract consumers of higher private cost, returns from the next batch of bike investment (which increases matching probability) will depend on the private cost of the next group of consumers. When the private cost distribution has a declining density (θ > 1), the same improvement in matching probability will persuade more consumers to search for a bike, and this effect increases with the number of consumers already in bike-sharing. It amounts to a market expanding effect, which speeds up the incentive in bike investment as long as the extra return in such investment exceeds the extra cost of investment. To summarize, the model points out three potential explanations for the market expansion effects shown in Section 5: (1) an increasing-return matching technology that generates large enough positive network effects (z = α + β > 1) but not large enough to dominate convex investment cost; (2) a declining density distribution of private cost that makes it easier to persuade the next batch of users to join bike-sharing (θ > 1); and (3) an awareness marketing campaign that enhances the multiplier in the matching technology (A) while keeping the same return-to-scale. The last explanation is easy to rule out because in our model the utilization rate is independent of A hence an entry-motivated increase in A could lead to growth in bike investment but cannot explain the increase in bike utilization rate. The other two explanations works together to explain higher investment and higher utilization rate, if their combination satisfies the conditions laid out in the model. Is there a way to further distinguish which of the two remaining explanations is the main driver of market expansion? We note that the two key parameters, z and θ, give us different comparative statistics. As shown in Figure 6 and Figure 5, the monopolist will invest more as z increases but invests less as θ increases. This is because higher θ means a steeper density

26 function of private cost hence for a given threshold of private cost (above which consumers will search for bike) the monopolist can only attract a smaller fraction of the population. This reduces the investment incentive in monopoly. On the contrary, as z increases, the matching technology becomes more efficient, thus the monopolist can attract a higher fraction of consumers by investing more. This will increase the monopolist’s investment incentive. The differential comparative statistics offers an opportunity to test whether the observed variations in market expansion is mainly driven by variations in z or in θ. In particular, some ofo-first cities had more bikes on the road than other ofo-first cities, before Mobike entered. As- suming these cities are comparable in everything else, if the difference in θ drives this difference, the cities with a higher pre-entry ofo investment should have a lower θ, which implies that they should experience smaller market expansion effects upon Mobike’s entry. On the contrary, if z drives the initial ofo investment, the first set of cities should have a higher zs, and therefore greater market expanding effects post entry. We will test this empirically.

7 Further Data Analysis Motivated by the Model

This section presents three sets of data analyses in light of the model. First, the model suggests that market expands because of the competition. In particular, the entrant’s bike investment enhances the matching rate and encourages more consumers to search for a generic bike once they start to search. This mechanism is most evident when ofo and Mobike bikes mingle together. Though we do not know exactly where Mobike puts its bikes within a city, the competition intensity between ofo and Mobike varies within ofo-First cities, because ofo had experienced a “campus period” when it restricted its operation within a college campus while Mobike always regards the whole city as the target market. Therefore, according to the model, the competition eﬀects should be weaker if ofo was still in the “campus period” when Mobike entered the city. To test this prediction, we decompose P ostEntryct into 1campus · P ostEntryct and (1 − 1campus) · P ostEntryct, and estimate their coeﬃcients separately. Table 6 shows the

OLS and IV results on log(Qct), pct and %F reect. Compared with the baseline results, we find that the market expanding effects are solely driven by the time when ofo expanded into the city. This finding confirms that the market expansion effects occur because ofo and Mobike compete head-in-head in the city. Our second set of analysis follows the comparative statistics implied by variations in z and θ. Greater market expansion effects could be driven by a higher θ, but higher θ implies greater reluctance to invest before the entry. This implies that we should observe greater market expansion in the ofo-First cities that had less bike investment before entry. In contrast, greater market expansion effects could be driven by a higher z and a higher z implies more investment before entry. This contrast leads us to include an interaction of pre-entry investment and the post entry dummy on the right hand side. Note that pre-entry investment only describes the cross- sectional variations across ofo First cities. To the extent that z and θ depends on pre-determined city attributes, pre-entry investment alone is absorbed by city fixed effects. As shown in Table 7, the positive price effect of entry is stronger when there had been more pre-entry investment in a city. The trade volume effect of entry goes the same direction, but the coefficient is only marginally significant. As articulated in the model, we need z and θ to work together to satisfy

27 the conditions for higher investment and higher utilization rate post entry. Table 7 does not reject this interdependence. Rather, it shows that the differential market expansion effects (in price and trade volume) is more likely due to variations in the extent of increasing return of matching (z) rather than variations in the distribution of consumers’ private cost (θ). The third analysis explores detailed geographic information in our dataset. Although our model abstracts away from geography within a city, one can imagine that residents at different parts of the city have different costs of using alternative transportation. For example, living on a street right next to a bus or subway station may make bike-sharing unnecessary. In contrast, living one kilometer away from the station could make bike-sharing much more attractive. Similarly, people working at the city-center, where it is easy to call a taxi or walk to a bus station, could be more reluctant to use bike-sharing than those working at a less convenient location. These variations give us a geographic interpretation of the distribution of the private cost c. As shown in the model, the monopolist first attracts those with the highest c and stops at a threshold c that makes the marginal consumer indifferent between searching for a bike and taking alternative transportation. If the monopolist knows where the high-c people are, it will place bikes near them. When the entrant enters the market, it will place its bikes near the next batch of consumers that have the highest c among those that have not chosen bike-sharing yet. To the extent that c might vary across people even if they live at the same location (for example some residents in an apartment may have private cars while others do not), the entrant could also place some bikes close to where the incumbent’s bikes have occupied before and enhance the probability of matching in the nearby area. Either way, entry could persuade people of lower c to join bike-sharing and these people may use either brand of bike depending on which is handy when they search. If they ride ofo bikes, these bikes will be available for the next rider at the destination. As a result, market expansion could geographically manifest in a network expansion of ofo bikes.

We use three variables to describe the geographic network of ofo bikes: #Gridsct describes the total number of unique grids covered by (the origin) of any ofo bike trips in a city-day;

Ginict describes how evenly distributed the origin of ofo bike trips is in the city-day; and a second version of Ginict is conditional on the grids that ofo has reached before Mobike’s entry. The last one depends on Mobike’s entry, so we can only compare it before and after the entry, without any control group. For the first two variables, we use the same DID specification as Equation (1). Table 8 reports the OLS and IV results for these three variables. They suggest that Mobike’s entry allows ofo bikes to reach more grids in the city and makes the ofo bikes distributed more evenly throughout the city. The network is also more evenly distributed within the grids that ofo has covered before the entry. Combined with other evidence, the geographic expansion confirms that Mobike’s entry attracts more users, enhances the reach of the ofo bike network, and boosts the average bike utilization rate.

8 Conclusion

Using proprietary data from a major bike-sharing firm, we document how entry affects the market performance of the incumbent. Since bike-sharing features positive network effects but

28 the market is city-specific, we have a rare opportunity to study competition with network effects. We find that the entrant expands the market, resulting in higher trade volume, higher price, higher bike investment, better bike utilization, and a wider, flatter network for the incumbent. However, the entrant also steals a significant fraction of the old users away from the incumbent, which in part justifies the entry decision. Our findings challenge the classical “winner-takes-all” concern in a market with network effects. According to that concern, positive network effects would enable the incumbent to become a natural monopoly and then abuse its monopoly power to the harm of consumers. In our context, entry creates positive spillovers on the incumbent, which helps the incumbent to better explore the positive network effects. This occurs for a couple of reasons: first, multi- homing consumers search for a generic bike, implying that one firm can motivate consumers to search but there is no guarantee that the search would lead to its own bike rather than the competitor’s bike. Second, the cost of investing and maintaining a diverse network of bikes is convex, thus it is more cost-efficient to free ride on the competitor’s investment than making all the investment on its own. In our model, the spillovers are mutual, which explains why the entrant finds it worthwhile to enter even if the incumbent has already operated in a market with positive network effects, and why the incumbent is willing to share the (expanded) market with the entrant. Furthermore, our work highlights the importance of the outside good in a network market. Since entry could generate market expansion, competition with the outside good is as important as within-market competition, for at least bike-sharing. These findings could have significant implications for policy makers, as they conduct merger reviews or consider entry policies in a market with positive network effects.

References

Alesina, Alberto, Stelios Michalopoulos, and Elias Papaioannou. “Ethnic Inequality.” Journal of Political Economy 124.2 (2016): 428-488. Angelucci, Charles, and Julia Cagé.“Newspapers in Times of Low Advertising Revenues.” Working Paper (2016). Armstrong, Mark. “Competition in Two-Sided Markets.” RAND Journal of Economics 37.3 (2006): 668-691. Autor, David H. “Outsourcing at Will: The Contribution of Unjust Dismissal Doctrine to the Growth of Employment Outsourcing.” Journal of Labor Economics 21.1 (2003): 1-42. Bleakley, Hoyt, and Jeffrey Lin. “Thick-market Effects and Churning in the Labor Market: Evidence from US Cities.” Journal of Urban Economics 72.2-3 (2012): 87-103. Bryan, Kevin A, and Joshua S. Gans. “A Theory of Multihoming in Rideshare Competition.” NBER Working Paper #24806. (2018)

29 Caillaud, Bernard, and Bruno Jullien. “Chicken & Egg: Competition among Intermediation Service Providers.” RAND Journal of Economics 34.2 (2003): 309-328. Chen, Yongmin, and Michael H. Riordan. “Priceincreasing Competition.” RAND Journal of Economics 39.4 (2008): 1042-1058. Chertok, M., A. Voukelatos, V. Sheppeard, et al. “Comparison of Air Pollution Exposure for Five Commuting Modes in Sydney-Car, Train, Bus, Bicycle and Walking.” Health Promotion Journal of Australia 15.1 (2004): 63-67. China Academy of Information and Communications Technology. “The Social and Economic Impacts of Bike-Sharing 2017.” (2018) DeMaio, Paul J. “Smart Bikes: Public Transportation for the 21st Century.” Transportation Quarterly 57.1 (2003): 9-11. DeMaio, Paul. “Bike-sharing: History, Impacts, Models of Provision, and Future.” Journal of Public Transportation 12.4 (2009): 3. DeMaio, Paul, and Jonathan Gifford. “Will Smart Bikes Succeed as Public Transportation in the United States?.” Journal of Public Transportation 7.2 (2004): 1. Dubé, Jean-Pierre H., Günter J. Hitsch, and Pradeep K. Chintagunta. “Tipping and Con- centration in Markets with Indirect Network Effects.” Marketing Science 29.2 (2010): 216-249. Duflo, Esther. “Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experiment.” American Economic Review 91.4 (2001): 795-813. Ellison, Glenn, Edward L. Glaeser, and William R. Kerr. “What Causes Industry Agglomera- tion? Evidence from Coagglomeration Patterns.” American Economic Review 100.3 (2010): 1195-1213. Farrell, Joseph, and Paul Klemperer. “Coordination and Lock-in: Competition with Switching Costs and Network Effects.” Handbook of industrial organization 3 (2007): 1967-2072. Farrell, Joseph, and Garth Saloner. “Installed Base and Compatibility: Innovation, Product Preannouncements, and Predation.” American Economic Review 76.5 (1986): 940-955. Gan, Li, and Qinghua Zhang. “The Thick Market Effect on Local Unemployment Rate Fluctuations.” Journal of Econometrics 133.1 (2006): 127-152. Gandal, Neil. “Hedonic Price Indexes for Spreadsheets and an Empirical Test for Network Externalities.” RAND Journal of Economics 25.1 (1994): 160-170. Gandal, Neil, Michael Kende, and Rafael Rob. “The Dynamics of Technological Adoption in Hardware/Software Systems: the Case of Compact Disc Players.” RAND Journal of Economics 31.1 (2000): 43-61. Gavazza, Alessandro. “The Role of Trading Frictions in Real Asset Markets.” American Economic Review 101.4 (2011): 1106-43. Halaburda, Hanna, and Yaron Yehezkel. “Platform Competition under Asymmetric Informa- tion.” American Economic Journal: Microeconomics 5.3 (2013): 22-68.

30 Hanson, Susan, and Perry Hanson. “Evaluating the Impact of Weather on Bicycle Use.” Transportation Research Record 629 (1977): 43-48. Hertel, O., M. Hvidberg, M. Ketzel, et al. “A Proper Choice of Route Significantly Reduces Air Pollution Exposure?A Study on Bicycle and Bus Trips in Urban Streets.” Science of the Total Environment 389.1 (2008): 58-70. Hopkinson, P. G., O. M. Carsten, and M. R. Tight. “Review of Literature on Pedestrian and Cyclist Route Choice Criteria.” (1989), available at https://trid.trb.org/view/1175944. Jacobson, Louis S., Robert J. LaLonde, and Daniel G. Sullivan. “Earnings Losses of Displaced Workers.” American Economic Review 83.4 (1993): 685-709. Kabra, Ashish, Elena Belavina, and Karan Girotra. “Bike-share Systems: Accessi- bility and Availability.” Chicago Booth Research Paper 15-04 (2016), available at https://ssrn.com/abstract=2555671. Kaiser, Ulrich, and Minjae Song. “Do Media Consumers Really Dislike Advertising? An Empirical Assessment of the Role of Advertising in Print Media Markets.” International Journal of Industrial Organization 27.2 (2009): 292-301. Kaiser, Ulrich, and Julian Wright. “Price Structure in Two-Sided Markets: Evidence from the Magazine Industry.” International Journal of Industrial Organization 24.1 (2006): 1-28. Katz, Michael L., and Carl Shapiro. “Network Externalities, Competition, and Compatibility.” American Economic Review 75.3 (1985): 424-440. Kim, Jin-Hyuk, Jeffrey Prince, and Calvin Qiu. “Indirect Network Effects and the Quality Di- mension: A Look at the Gaming Industry.” International Journal of Industrial Organization 37 (2014): 99-108.

Liang, Xuan, et al. “PM2.5 Data Reliability, Consistency, and Air Quality Assessment in Five Chinese Cities.” Journal of Geophysical Research: Atmospheres 121.17 (2016): 10-220. Mankiw, N. Gregory, and Michael D. Whinston. “Free Entry and Social Inefficiency.” RAND Journal of Economics (1986): 48-58. Murry, Charles, and Yiyi Zhou. “Consumer Search and Automobile Dealer Co-Location.” Working Paper (2018). Nankervis, Max. “The Effect of Weather and Climate on Bicycle Commuting.” Transportation Research Part A: Policy and Practice 33.6 (1999): 417-431. O’Mahony, Eoin, and David B. Shmoys. “Data Analysis and Optimization for (Citi) Bike- sharing.” AAAI (2015). Pan, Ling, et al. “Rebalancing Dockless Bike-sharing Systems.” arXiv preprint arXiv:1802.04592 (2018). Park, Sangin. “Quantitative Analysis of Network Externalities in Competing Technologies: The VCR Case.” Review of Economics and Statistics 86.4 (2004): 937-945. Petrongolo, Barbara, and Christopher Pissarides. “Scale Effects in Markets with Search.” The Economic Journal 116.508 (2006): 21-44.

31 Rosenthal, Robert W. “A Model in which an Increase in the Number of Sellers Leads to a Higher Price.” Econometrica: Journal of the Econometric Society (1980): 1575-1579. Rysman, Marc. “Competition between Networks: A Study of the Market for Yellow Pages.” Review of Economic Studies 71.2 (2004): 483-512. Saloner, Garth and Andrea Shepard. “Adoption of Technologies with Network Effects: An Empirical Examination of the Adoption of Automated Teller Machines.” RAND Journal of Economics 26.3 (1995): 479-501. Seamans, Robert and Feng Zhu. “Responses to Entry in Multi-Sided Markets: The Impact of Craigslist on Local Newspapers.” Management Science 60.2 (2013): 476-493. Stahl, Dale O. “Oligopolistic Pricing with Sequential Consumer Search.” American Economic Review 79.4 (1989): 700-712. Vicente-Serrano, Sergio M., M. Angel Saz-Sánchez, and José M. Cuadrat. “Comparative Analysis of Interpolation Methods in the Middle Ebro Valley (Spain): Application to Annual Precipitation and Temperature.” Climate Research 24.2 (2003): 161-180. Vitorino, Maria Ana. “Empirical Entry Games with Complementarities: An Application to the Shopping Center Industry.” Journal of Marketing Research 49.2 (2012): 175-191. Zhang, Yongping, and Zhifu Mi. “Environmental Benefits of Bike-sharing: A Big Data-based Analysis.” Applied Energy 220 (2018): 296-301. Zhang, Lihong, et al. “Sustainable Bike-sharing Systems: Characteristics and Commonalities across Cases in Urban China.” Journal of Cleaner Production 97 (2015): 124-133. Zheng, Fanyin, et al. “Customer Preference and Station Network in the London Bike Share System.” Columbia Business School Research Paper 18-20 (2018), available at https://ssrn.com/abstract=3121791.

32 Table 1: Data Summary (mean of key outcomes are masked by “NA” for conﬁdentiality) Sample Full Sample ofo First ofo Alone Mobike First Variables Mean Std. Dev Mean Std. Dev Mean Std. Dev Mean Std. Dev Panel A City-Day Level Variables Sample Size 19631 13560 2633 3438 Dummy for Post-Entry Status 0.616 0.486 0.639 0.48 0 0 1 0 Log (Trip Volume) NA 2.124 NA 2.074 NA 1.386 NA 1.968 Average Revenue per Trip (RMB) NA 0.207 NA 0.195 NA 0.229 NA 0.23 Percent of Free Trips (0-100) NA 22.774 NA 23.716 NA 19.869 NA 20.276 Log (# of New Users) NA 1.997 NA 1.974 NA 1.789 NA 1.648 Percent of Active Old Users NA 13.668 NA 12.955 NA 16.471 NA 13.905 Average # of Trips per Old User NA 0.388 NA 0.41 NA 0.351 NA 0.305 Log (# of Grids Covered by ofo) 5.469 1.21 5.539 1.265 4.709 0.862 5.777 0.959 33 Gini Coverage Index 0.864 0.09 0.882 0.086 0.788 0.092 0.852 0.067 Dummy for ofo Operation within Campus 0.163 0.369 0.22 0.415 0 0 0.061 0.24 Speed of Wind 2.677 0.883 2.661 0.901 2.679 0.861 2.74 0.823 Temperature 21.276 7.69 20.167 8.164 23.515 5.708 23.935 5.838 Precipitation 0.171 0.486 0.152 0.455 0.208 0.519 0.219 0.567 Relative Humility 73.831 16.31 72.662 16.786 74.259 16.763 78.115 12.99 AQI (Air Quality Index) 84.196 47.688 87.795 51.345 77.956 37.909 74.782 36.304 Panel B City Level Variables Sample Size 104 59 23 22 Logarithmic Population (10,000) 6.101 0.632 6.196 0.558 5.814 0.818 6.143 0.535 GDP per Capita (10,000 RMB) 6.699 3.356 6.94 3.254 5.981 3.118 6.804 3.885 Number of Taxis 5076.103 6903.375 6351.334 6265.299 2040.783 1951.386 4829.455 10325.5 Number of Buses 3029.936 4474.697 3774.835 4871.471 942.235 730.883 3214.847 5073.091 Road Surface (10,000 Square Meters) 3281.349 3189.038 3951.99 3476.455 1627.596 1267.843 3211.735 3248.628 Number of Mobile Phone Users (10,000) 688.125 585.702 806.576 603.333 366.957 217.516 706.227 689.128 Number of Internet Households (10,000) 142.173 159.78 170.593 184.558 70.609 44.449 140.773 145.571 Average Gradient () 458.766 570.863 447.244 547.904 596.891 745.032 345.262 391.15 Table 2: Competition Eﬀects on Usage Volume and Price

(1) (2) (3) (4) (5) (6) (7) Panel A Dependent Variable Log (Trip Volume) PostEntry 0.370* 0.439** 0.535*** 0.491** 0.346** 0.402** 0.408** (0.211) (0.181) (0.199) (0.207) (0.166) (0.181) (0.185) Within Adjusted R2 0.104 0.233 0.106 0.111 0.241 0.12 0.117 Panel B Dependent Variable Average Revenue per Trip PostEntry 0.029*** 0.027*** 0.030*** 0.035*** 0.031*** 0.031*** 0.041*** (0.009) (0.008) (0.010) (0.011) (0.008) (0.010) (0.011) Within Adjusted R2 0.075 0.122 0.049 0.057 0.123 0.049 0.059 Panel C Dependent Variable Percent of Free Trips (0-100) PostEntry -2.288** -2.311** -3.132** -3.589** -2.170** -2.717** -3.695*** (1.131) (1.117) (1.487) (1.563) (0.971) (1.287) (1.399) Within Adjusted R2 0.085 0.14 0.073 0.07 0.14 0.074 0.07 Dummy for Operation within Campus YES YES YES YES YES YES YES Weather Condition YES YES YES YES YES YES YES Air Quality YES YES YES YES YES YES YES City Fixed Effects YES YES YES YES YES YES YES Calendar Date Fixed Effects YES YES YES YES YES YES YES Relative Day Fixed Effects YES YES YES YES YES YES YES Duopoly Group Trend YES YES YES Linear Time Trend YES YES City AttributesDate Fixed Effects YES YES City AttributesDay Fixed Effects YES YES Number of Clusters 82 82 82 82 82 82 82 Number of Observations 16193 16193 16193 16193 16193 16193 16193

Notes: Column 1 only controls for city fixed effects and time fixed effects. Column 2 adds the interaction between predetermined city attributes and a third-order polynomial function of the relative days since ofos entry. Column 3 and 4 interact the city attributes with calendar date fixed effects and relative day fixed effects respectively. Column 5-7 further include the linear time trend specific to the ofo First cites. The specification of Column 7 is taken as benchmark settings in the following analyses. Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level.

34 Table 3: 2SLS Estimates

(1) (2) (3) (4) Average Revenue Percent of Dependent Variables PostEntry Log (Trip Volume) per Trip Free Trips Models First-Stage 2SLS 2SLS 2SLS Predicted PostEntry 0.949*** (0.011) PostEntry 0.478** 0.045*** -3.999*** (0.199) (0.012) (1.493) Dummy for Operation within Campus YES YES YES YES Weather Condition YES YES YES YES Air Quality YES YES YES YES City Fixed Effects YES YES YES YES Calendar Date Fixed Effects YES YES YES YES Relative Day Fixed Effects YES YES YES YES Duopoly Group Trend YES YES YES YES City AttributesDay Fixed Effects YES YES YES YES Kleibergen-Paap F Test 8000.251 / / / Number of Clusters 82 82 82 82 Number of Observations 16193 16193 16193 16193 Notes: The instrument variable Predicted PostEntry is derived from a duration model which treats the time span between Mobike entry dates and November 1, 2015 as “survival time” and uses city attributes and VC finance of Mobike as regressors. We assume that the baseline hazard follows Weibull distribution. Column 1 reports the first-stage with the Kleibergen- Paap F test larger than 8000. Column 2-4 show 2SLS estimates under the benchmark settings, which are similar to baseline results in both significance and magnitude. Further robustness checks of starting date choice and the assumption of baseline hazards are reported in Appendix Table A3. Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level.

35 Table 4: Competition Eﬀects on Bike Utilization Rate and Bike Investment

(1) (2) (3) (4) Dependent Variables Log (Bike Utilization Rate) Bike Investment Models OLS 2SLS OLS 2SLS PostEntry 0.392** 0.457** (0.185) (0.198) Percent of Duopoly Days (1-100) 57.526** 58.384** (27.789) (27.391) Dummy for (Percent of) Operation within Campus YES YES YES YES Weather Condition YES YES YES YES Air Quality YES YES YES YES City Fixed Effects YES YES YES YES Calendar Date(Month) Fixed Effects YES YES YES YES Relative Day(Month) Fixed Effects YES YES YES YES Duopoly Group Trend YES YES YES YES City AttributesDay(Month) Fixed Effects YES YES YES YES Within Adjusted R2 0.092 / 0.016 / Number of Clusters 79 79 79 79 Number of Observations 15770 15770 616 616 Notes: Every two columns under the same outcome variable report OLS and 2SLS estimates separately which adopt the benchmark settings in Table 2 Column 7. Because of the lumpi- ness bike investment, we aggregate the investment data to month level, redefine PostEntry as the percent of days that Mobike operates in the city c and month m, reconstruct all weather and air variables as monthly average, and control for monthly time fixed effects instead of daily fixed effects. Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level.

36 Table 5: Market Expanding vs. Market Stealing Eﬀects

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Percent of Average # of Trips Average Revenue per Average Revenue per Dependent Variables Log (# of New Users) Active Old Users per Old User Trip (New Users) Trip (Old Users) Models OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS PostEntry 0.652*** 0.735*** -4.126*** -4.353*** -0.005 -0.003 0.029*** 0.030*** 0.032*** 0.034*** (0.228) (0.243) (1.446) (1.551) (0.036) (0.039) -0.007 -0.008 -0.008 -0.009 Dummy for Operation within Campus YES YES YES YES YES YES YES YES YES YES Weather Condition YES YES YES YES YES YES YES YES YES YES Air Quality YES YES YES YES YES YES YES YES YES YES City Fixed Effects YES YES YES YES YES YES YES YES YES YES Calendar Date Fixed Effects YES YES YES YES YES YES YES YES YES YES Relative Day Fixed Effects YES YES YES YES YES YES YES YES YES YES Duopoly Group Trend YES YES YES YES YES YES YES YES YES YES City AttributesDay FE YES YES YES YES YES YES YES YES YES YES Within Adjusted R2 0.265 / 0.075 / 0.006 / 0.007 / 0.041 / Number of Clusters 82 82 82 82 82 82 82 82 82 82 Number of Observations 16193 16193 16193 16193 16193 16193 16193 16193 16193 16193

Notes: Every two columns under the same outcome variable report OLS and 2SLS estimates separately which adopt the benchmark settings in Table 2 Column 7. Standard errors are in parentheses and clustered at the city level. *** Denotes signiﬁcance at the 1%, ** 5%, and * 10% level. Table 6: Placebo Test Using ofo Campus Period

(1) (2) (3) (4) (5) (6) Dependent Variables Log(Trip Volume) Average Revenue per Trip Percent of Free Trips Models OLS 2SLS OLS 2SLS OLS 2SLS PostEntryDummy for Operation within Campus -0.371 -0.211 0.000 0.016 0.367 -0.758 (0.298) (0.293) (0.027) (0.028) (4.100) (4.068) PostEntryDummy for Operation in the Whole City 0.618*** 0.673*** 0.060*** 0.063*** -5.451*** -5.688*** (0.190) (0.200) (0.012) (0.013) (1.505) (1.588) Dummy for Operation within Campus NO NO NO NO NO NO Weather Condition YES YES YES YES YES YES Air Quality YES YES YES YES YES YES 38 City Fixed Effects YES YES YES YES YES YES Calendar Date Fixed Effects YES YES YES YES YES YES Relative Day Fixed Effects YES YES YES YES YES YES Duopoly Group Trend YES YES YES YES YES YES City AttributesDay Fixed Effects YES YES YES YES YES YES Within Adjusted R2 0.107 / 0.042 / 0.056 / Number of Clusters 82 82 82 82 82 82 Number of Observations 16193 16193 16193 16193 16193 16193

Notes: The key independent variable P ostEntryct is decomposed into 1campus · P ostEntryct and (1 − 1campus) · P ostEntryct, where 1campus is the dummy for operation within campus in benchmark settings. Every two columns under the same outcome variable report OLS and 2SLS estimates separately. Column 1-6 point to the common conclusion that the market expanding effects emerge when ofo expanded to the whole city while absent during “campus period.” Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level. Table 7: Competition Effect Heterogeneity to Pre-Entry Bike Investment

(1) (2) (3) (4) (5) (6) Dependent Variables Log(Trip Volume) Average Revenue per Trip Percent of Free Trips Models OLS 2SLS OLS 2SLS OLS 2SLS PostEntry 0.121 0.198 0.017 0.021 -1.429 -1.765 (0.216) (0.233) (0.013) (0.014) (1.500) (1.625) PostEntryPre-Entry Bike Investment 0.036* 0.034 0.003*** 0.003*** -0.296* -0.288* (0.021) (0.021) (0.001) (0.001) (0.149) (0.149) Dummy for Operation within Campus YES YES YES YES YES YES Weather Condition YES YES YES YES YES YES Air Quality YES YES YES YES YES YES 39 City Fixed Effects YES YES YES YES YES YES Calendar Date Fixed Effects YES YES YES YES YES YES Relative Day Fixed Effects YES YES YES YES YES YES Duopoly Group Trend YES YES YES YES YES YES City AttributesDay Fixed Effects YES YES YES YES YES YES Within Adjusted R2 0.129 / 0.068 / 0.077 / Number of Clusters 81 81 81 81 81 81 Number of Observations 16140 16140 16140 16140 16140 16140 Notes: The Pre-Entry Bike Investment is calculated as the average number of accumulative bike investment over 10 days before Mobikes entry. If the time gap between the entry of ofo and Mobike is shorter than 10 days, this index is still constructed over all the gap days thought from less observations. To make the magnitude of coefficients suitable for understanding , we divide the number of pre-entry investment by 1,000. Bike investment data is missing for one of ofo First cities and the number of clusters thus decreases to 81. Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level. Table 8: Competition Effects on Geographical Reach of Bike-Sharing Network

(1) (2) (3) (4) (5) (6) Log (# of Grids Gini Coverage Index Dependent Variables Gini Coverage Index covered by ofo) of Pre-Entry Grids Models OLS 2SLS OLS OLS 2SLS OLS PostEntry 0.195** 0.225** -0.035*** -0.038*** -0.027** -0.031*** (0.081) (0.086) (0.007) (0.008) (0.010) (0.011) Dummy for Operation within Campus YES YES YES YES YES YES Weather Condition YES YES YES YES YES YES Air Quality YES YES YES YES YES YES 40 City Fixed Effects YES YES YES YES YES YES Calendar Date Fixed Effects YES YES YES YES YES YES Relative Day Fixed Effects YES YES YES YES YES YES City AttributesDay Fixed Effects YES YES YES YES YES YES Duopoly Group Trend YES YES YES YES NO NO Within Adjusted R2 0.202 / 0.112 / 0.041 / Number of Clusters 82 82 82 82 59 59 Number of Observations 16193 16193 16193 16193 13560 13560 Notes: “Pre-Entry Grids” could not be defined for ofo Alone group cities, which are excluded in Column 5 and 6 and the number of clusters decreases to 59 (i.e., the number of ofo Frist group cities). Every two columns under the same outcome variable report OLS and 2SLS estimates separately which adopt the benchmark settings in Table 2 Column 7. Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level. Figure 1: Expension Process of ofo and Mobike

(a) September 30, 2016 (b) December 31, 2016

Notes: This ﬁgure depicts the expansion process of ofo and Mobike in our sample cities. Beijing and the 6 cities without detailed entry sequence are excluded. The base map of China comes from Resource and Environment Data Cloud Platform (http://www.resdc.cn).

41 Figure 2: Test of Common Pre-Trend Assumption

(a) Log(Trip Volume)

(b) Average Revenue per Trip

(d) Log(Bike Utilization Rate)

(e) Bike Investment

k=21 Notes: Point estimates of {λ−k}k=2 in Equation (2) as well as corresponding 95% confidence intervals are plotted with relative days before Mobike’s entry on the horizontal axis. The day before Mobike’s entry is omitted as base and days more than 3 weeks before the entry are all counted as 21. Similar notations apply to the last panel with time unit changed into month. All the coefficients are indistinguishable from 0 even at the 10% significance level, which implies that ofo Alone and ofo First cities follow similar pre-entry trends. All the other controls are the same as Table 2 Column 7. Standard errors are clustered at the city level.

42 Figure 3: Falsiﬁcation Test of Forwards of PostEntry

(a) Log(Trip Volume)

(b) Average Revenue per Trip

(d) Log(Bike Utilization Rate)

(e) Bike Investment

Notes: We restrict the sample to ofo Alone cities and pre-entry observations of ofo First cities, and generate false entry on 1,2,,7 days before the publicly announced Mobike entry. Point estimates of the false entry as well as 95% conﬁdence intervals are depicted together with the baseline estimates from Table 2 Column 7 plotted on the very right. Similar notations apply to the last panel with time unit changed into month. Standard errors are clustered at the city level.

43 Figure 4: The graph plots how the parameter values α and β aﬀect the comparison of rd and rm when we ﬁx θ = γ = 2.

1 upper-bound 0.9 of -

0.8

rd>rm 0.7

0.6 -

0.5 rd

0.4

0.3

0.2

0.1 0.5 0.6 0.7 0.8 0.9 1 ,

44 Figure 5: The graphs plot how the equilibrium prices, investments, trade volumes and utilization rates change with z = α + β in both monopoly and duopoly cases. 3.5 0.8 d pd v m pm v 0.75 3

0.7 2.5

0.65 v

p 2

0.6

1.5 0.55

1 0.5

0.45 0.5 1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6 z z (a) Price as a function of z (b) Investment as a function of z 4.5 1.3 d tv rd m m 4 tv 1.2 r

1.1 3.5

1 3 0.9 2.5 0.8 2

total volume 0.7 utilization rate 1.5 0.6 1 0.5

0.5 0.4

0 0.3 1 1.1 1.2 1.3 1.4 1.5 1.6 1 1.1 1.2 1.3 1.4 1.5 1.6 z z (c) Trade volume as a function of z (d) Utilization rate as a function of z

45 Figure 6: The graphs plot how the equilibrium prices, investments, trade volumes and utilization rates change with θ in both monopoly and duopoly cases. 1 1.5 d pd v m pm 1.4 v 0.9 1.3

0.8 1.2

1.1 0.7 v

p 1 0.6 0.9

0.8 0.5

0.7 0.4 0.6

0.3 0.5 0 0.5 1 1.5 2 0 0.5 1 1.5 2 3 3 (a) Price as a function of θ (b) Investment as a function of θ

1.2 0.75 rd tvd rm tvm 0.7 1 0.65

0.6 0.8

0.55

0.6 0.5

trade volume 0.45 utilization rate 0.4 0.4

0.35 0.2 0.3

0 0.25 0 0.5 1 1.5 2 0 0.5 1 1.5 2 3 3 (c) Trade volume as a function of θ (d) Utilization rate as a function of θ

46 Appendix

Table A1: List of Cities Administrative City Name ofo Entry Date Mobike Entry Date Group Area Code Tianjin 120000 27-Aug-16 12-Feb-17 ofo First Shijiazhuang 130100 31-Aug-16 6-Mar-17 ofo First Tangshan 130200 1-Apr-17 17-Apr-17 ofo First Qinhuangdao 130300 28-Apr-17 12-Jun-17 ofo First Handan 130400 14-Apr-17 6-May-17 ofo First Baoding 130600 9-Mar-17 19-Jun-17 ofo First Langfang 131000 20-Apr-17 17-May-17 ofo First Taiyuan 140100 17-Aug-16 14-May-17 ofo First Datong 140200 3-Mar-17 27-Jun-17 ofo First Jinzhong 140700 6-May-17 17-May-17 ofo First Xinzhou 140900 10-Jul-17 / ofo Alone Hohhot 150100 1-May-17 / ofo Alone Wuhai 150300 30-Jun-17 / ofo Alone Erdos 150600 9-Jun-17 8-May-17 Mobike First Shenyang 210100 8-May-17 17-May-17 ofo First Dalian 210200 26-Jun-17 16-Apr-17 Mobike First Shanghai 310000 9-May-16 22-Apr-16 Mobike First Nanjing 320100 14-Jun-16 12-Jan-17 ofo First Wuxi 320200 2-Mar-17 3-Mar-17 ofo First Suzhou 320500 15-Jan-17 18-Jun-17 ofo First Nantong 320600 29-Apr-17 / ofo Alone Yangzhou 321000 20-Apr-17 9-Mar-17 Mobike First Zhenjiang 321100 28-Apr-17 / ofo Alone Hangzhou 330100 12-Sep-16 16-Apr-17 ofo First Ningbo 330200 14-Jan-17 6-Dec-16 Mobike First Wenzhou 330300 14-May-17 8-Apr-17 Mobike First Jiaxing 330400 6-Apr-17 27-Apr-17 ofo First Jinhua 330700 31-Mar-17 20-May-17 ofo First Taizhou 331000 18-May-17 1-Jul-17 ofo First Hefei 340100 24-Aug-16 13-Feb-17 ofo First Wuhu 340200 16-Mar-17 26-Mar-17 ofo First Maanshan 340500 28-Dec-16 11-May-17 ofo First Anqing 340800 6-Dec-16 / ofo Alone Fuzhou 350100 19-Aug-16 7-Feb-17 ofo First Xiamen 350200 17-Dec-16 20-Dec-16 ofo First Quanzhou 350500 14-Mar-17 8-Mar-17 Mobike First Zhangzhou 350600 13-Mar-17 9-Mar-17 Mobike First Ningde 350900 25-Apr-17 / ofo Alone

47 Nanchang 360100 20-Aug-16 24-Feb-17 ofo First Jiujiang 360400 20-Apr-17 20-May-17 ofo First Ganzhou 360700 20-Apr-17 16-Jun-17 ofo First Shangrao 361100 14-May-17 / ofo Alone Jinan 370100 29-Aug-16 25-Jan-17 ofo First Qingdao 370200 21-Feb-17 7-May-17 ofo First Zibo 370300 3-Apr-17 / ofo Alone Zaozhuang 370400 29-Jun-17 17-May-17 Mobike First Yantai 370600 5-May-17 / ofo Alone Weifang 370700 28-Apr-17 / ofo Alone Jining 370800 17-Jun-17 17-May-17 Mobike First Tai’an 370900 10-Apr-17 23-May-17 ofo First Weihai 371000 25-Apr-17 7-May-17 ofo First Rizhao 371100 29-Apr-17 19-Mar-17 Mobike First Dezhou 371400 23-May-17 27-Apr-17 Mobike First Zhengzhou 410100 11-Aug-16 6-Mar-17 ofo First Kaifeng 410200 17-May-17 17-May-17 ofo First Luoyang 410300 20-Apr-17 10-Apr-17 Mobike First Puyang 410900 22-Jul-17 11-Aug-17 ofo First Xuchang 411000 4-Jun-17 / ofo Alone Sanmenxia 411200 19-Jun-17 / ofo Alone Wuhan 420100 18-Apr-16 29-Dec-16 ofo First Shiyan 420300 19-Aug-17 / ofo Alone Yichang 420500 9-Apr-17 7-Apr-17 Mobike First Xiangyang 420600 2-Apr-17 1-May-17 ofo First Ezhou 420700 16-May-17 16-Jul-17 ofo First Xiaogan 420900 10-May-17 / ofo Alone Huanggang 421100 15-May-17 25-Aug-17 ofo First Xianning 421200 6-Jun-17 12-Jun-17 ofo First Changsha 430100 26-Aug-16 14-Feb-17 ofo First Zhuzhou 430200 24-Apr-17 / ofo Alone Xiangtan 430300 24-Apr-17 / ofo Alone Guangzhou 440100 8-Jun-16 27-Sep-16 ofo First Shaoguan 440200 1-Jun-17 / ofo Alone Shenzhen 440300 11-Sep-16 16-Oct-16 ofo First Zhuhai 440400 20-Oct-16 21-Jan-17 ofo First Shantou 440500 12-Apr-17 19-Feb-17 Mobike First Jiangmen 440700 10-Apr-17 27-Mar-17 Mobike First Heyuan 441600 9-Jun-17 / ofo Alone Dongguan 441900 24-Feb-17 13-Jan-17 Mobike First Zhongshan 442000 7-Apr-17 16-Jun-17 ofo First Jieyang 445200 17-Apr-17 / ofo Alone Nanning 450100 7-Sep-16 21-Feb-17 ofo First

48 Guilin 450300 1-Mar-17 30-May-17 ofo First Haikou 460100 28-Feb-17 17-Feb-17 Mobike First Chengdu 510100 22-Aug-16 16-Nov-16 ofo First Deyang 510600 22-Apr-17 9-Mar-17 Mobike First Mianyang 510700 17-Mar-17 6-Mar-17 Mobike First Leshan 511100 10-May-17 17-May-17 ofo First Nanchong 511300 8-May-17 17-May-17 ofo First Meishan 511400 8-Jul-17 23-Jun-17 Mobike First Ziyang 512000 1-Jun-17 23-May-17 Mobike First Guiyang 520100 6-Mar-17 9-Apr-17 ofo First Liupanshui 520200 6-May-17 / ofo Alone Zunyi 520300 27-Apr-17 21-May-17 ofo First Kunming 530100 27-Aug-16 8-Jan-17 ofo First Xi’an 610100 27-May-16 19-Feb-17 ofo First Xianyang 610400 29-Apr-17 17-May-17 ofo First Weinan 610500 20-May-17 21-May-17 ofo First Yan’an 610600 22-May-17 16-Aug-17 ofo First Yulin 610800 23-May-17 3-Aug-17 ofo First Lanzhou 620100 25-Aug-16 10-Jul-17 ofo First Xining 630100 8-May-17 / ofo Alone Yinchuan 640100 25-Apr-17 25-Apr-17 ofo First Urumqi 650100 5-Jul-17 7-Jul-17 ofo First Karamay 650200 22-Aug-17 / ofo Alone

Notes: This list only includes cities in our ﬁnal sample. Beijing and the 6 cities without detailed entry sequence are excluded. Administrative Area Code is a unique number to identify administrative area, which is issued by the China central government. / means that entry dates are missing for ofo Alone cities.

49 Table A2: Robustness Check of Diﬀerent Subsamples

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Dependent Variables Log (Trip Volume) Average Revenue per Trip Percent of Free Trips Subsamples ofo Alone Mobike First ofo Alone Mobike First ofo Alone Mobike First Excluded Included Excluded Included Excluded Included Models OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS PostEntry 0.401** 0.473** 0.373* 0.428** 0.044*** 0.048*** 0.039*** 0.042*** -3.245** -3.510** -3.600*** -3.804*** (0.192) (0.206) (0.190) (0.202) (0.013) (0.014) (0.011) (0.012) (1.307) (1.387) (1.297) (1.376) Dummy for Operation within Campus YES YES YES YES YES YES YES YES YES YES YES YES Weather Condition YES YES YES YES YES YES YES YES YES YES YES YES

50 Air Quality YES YES YES YES YES YES YES YES YES YES YES YES City Fixed Effects YES YES YES YES YES YES YES YES YES YES YES YES Calendar Date Fixed Effects YES YES YES YES YES YES YES YES YES YES YES YES Relative Day Fixed Effects YES YES YES YES YES YES YES YES YES YES YES YES Duopoly Group Trend NO NO YES YES NO NO YES YES NO NO YES YES City AttributesDay Fixed Effects YES YES YES YES YES YES YES YES YES YES YES YES Within Adjusted R2 0.103 / 0.133 / 0.074 / 0.053 / 0.079 / 0.061 / Number of Clusters 59 59 104 104 59 59 104 104 59 59 104 104 Number of Observations 13560 13560 19631 19631 13560 13560 19631 19631 13560 13560 19631 19631 Notes: This table further examines the robustness of results in Table 2 and 3. Column 1,2,5,6,9 and 10 drop ofo Alone cities and re-estimate the coefficients under the benchmark specification, resulting from the concern that our list of controls could not fully guarantee the comparability between ofo First and ofo Alone cities. The other columns include the Mobike First group which is equivalent to the “always-treated” group in the context of DID framework and make full use of the data sample. Standard errors are in parentheses and clustered at the city level. *** Denotes significance at the 1%, ** 5%, and * 10% level. Table A2: Robustness Check of Different Subsamples (Continued)

(13) (14) (15) (16) (17) (18) (19) (20) Dependent Variables Log (Bike Utilization Rate) Bike Investment Subsamples ofo Alone Mobike First ofo Alone Mobike First Excluded Included Excluded Included Models OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS PostEntry 0.397** 0.464** 0.357* 0.409** -0.193 -0.206 -0.192 -0.202 Percent of Duopoly Days (1-100) 51.669* 52.144* 40.842* 43.024* -30.894 -30.413 -21.713 -22.03 Dummy for (Percent of) Operation within Campus YES YES YES YES YES YES YES YES

51 Weather Condition YES YES YES YES YES YES YES YES Air Quality YES YES YES YES YES YES YES YES City Fixed Effects YES YES YES YES YES YES YES YES Calendar Date(Month) Fixed Effects YES YES YES YES YES YES YES YES Relative Day(Month) Fixed Effects YES YES YES YES YES YES YES YES Duopoly Group Trend NO NO YES YES NO NO YES YES City AttributesDay(Month) Fixed Effects YES YES YES YES YES YES YES YES Within Adjusted R2 0.103 / 0.109 / 0.01 / 0.002 / Number of Clusters 58 58 101 101 58 58 101 101 Number of Observations 13507 13507 19208 19208 515 515 754 754 Table A3: Robustness Check of 2SLS Estimates

(1) (2) (3) (4) (5) Starting Dates 12/1/15 1/1/16 2/1/16 3/1/16 4/1/16 Distribution Panel A Dependent Variables Log (Trip Volume) Weibull 0.484** 0.480** 0.487** 0.498** 0.505** (0.201) (0.202) (0.204) (0.205) (0.206) Loglogistic 0.450** 0.456** 0.464** 0.469** 0.479** (0.198) (0.199) (0.200) (0.201) (0.205) Lognormal 0.456** 0.459** 0.461** 0.469** 0.477** (0.199) (0.201) (0.202) (0.204) (0.207) Panel B Dependent Variables Average Revenue per Trip Weibull 0.045*** 0.045*** 0.046*** 0.046*** 0.046*** (0.012) (0.012) (0.012) (0.013) (0.013) Loglogistic 0.043*** 0.044*** 0.044*** 0.045*** 0.046*** (0.012) (0.012) (0.012) (0.012) (0.013) Lognormal 0.044*** 0.043*** 0.043*** 0.043*** 0.043*** (0.012) (0.012) (0.012) (0.013) (0.013) Panel C Dependent Variables Percent of Free Trips Weibull -4.023*** -4.031*** -4.101*** -4.153*** -4.158*** (1.502) (1.508) (1.522) (1.534) (1.537) Loglogistic -3.796** -3.804** -3.889** -3.962** -4.017** (1.496) (1.506) (1.508) (1.519) (1.542) Lognormal -3.872** -3.832** -3.804** -3.851** -3.837** (1.489) (1.505) (1.517) (1.533) (1.554) Notes: The five panels experiment with instrument variables constructed from duration models that use December 1, 2015, January 1, 2016, February 1, 2016, March 1, 2016 and April 1, 2016 as starting dates of Mobike, under different assumptions for the functional form of baseline hazard (that is, Weibull, log-log and log-normal distributions). For each outcome variable, there are 53 = 15 estimates of β. This table provides further support to Table 3 and Table 4 in the sense that results are not driven by the choice of starting dates or distribution function. Standard errors are in parentheses and clustered at the city level. ***Denotes significance at the 1%, ** 5%, and * 10% level.

52 Table A3: Robustness Check of 2SLS Estimates (Continued)

(1) (2) (3) (4) (5) Starting Dates 12/1/15 1/1/16 2/1/16 3/1/16 4/1/16 Distribution Panel D Dependent Variable Log (Bike Utilization Rate) Weibull 0.461** 0.457** 0.463** 0.473** 0.480** (0.200) (0.200) (0.202) (0.203) (0.204) Loglogistic 0.428** 0.434** 0.440** 0.445** 0.454** (0.196) (0.198) (0.198) (0.200) (0.202) Lognormal 0.433** 0.436** 0.437** 0.444** 0.452** (0.198) (0.200) (0.200) (0.202) (0.205) Panel E Dependent Variable Bike Investment Weibull 52.342* 52.066* 51.285* 51.632* 51.866* (30.428) (30.571) (30.217) (30.292) (30.321) Loglogistic 47.565 47.501 47.696 47.436 46.338 (29.457) (29.515) (29.486) (29.285) 46.338 Lognormal 48.052 47.765 46.609 47.315 46.894 (29.610) (29.754) (29.490) (29.435) (29.289)

53 Table A4: Competition Eﬀects on Usage Volume within Pre-Entry & Non-Campus Grids

(1) (2) (3) (4) Dependent Variable Log (Trip Volume) Models OLS 2SLS OLS 2SLS PostEntry 0.386** 0.451** 0.440** 0.490** (0.190) (0.205) (0.194) (0.210) Dummy for Operation within Campus YES YES YES YES Weather Condition YES YES YES YES Air Quality YES YES YES YES City Fixed Effects YES YES YES YES Calendar Date Fixed Effects YES YES YES YES Relative Day Fixed Effects YES YES YES YES City AttributesDay Fixed Effects YES YES YES YES Duopoly Group Trend YES YES YES YES Within Adjusted R2 0.101 / 0.045 / Number of Clusters 59 59 53 53 Number of Observations 13560 13560 9170 9170 Notes: To investigate whether the booming usage is mainly driven by expansion to new grids, we first restrict to grids which are already covered by ofo before Mobikes entry and re- compute usage volume. ofo Alone and Mobike First cities are thus excluded and estimates are reported in Column 1 and 2. To eliminate potential effects from campus, we further restrict to non-campus grids among those old grids which are employed in the regression for Column 1 and 2. Please note that some cities of ofo First group have been covered by ofo completely during the campus period, i.e., ofo does not strictly enforce the within- campus strategy. We could not define non-campus grids for them and the number of clusters decreases to 53 in Column 3 and 4. Standard errors are in parentheses and clustered at the city level. ***Denotes significance at the 1%, ** 5%, and * 10% level.